Month: August 2024

Retrieval augmented generation (RAG) allows large language models (LLMs) to answer queries related to the data the models have not seen during training. In my previous article, I explained how to develop RAG systems using the Claude 3.5 Sonnet model.

However, RAG systems only answer queries about the data stored in the vector database. For example, you have a RAG system that answers queries related to financial documents in your database. If you ask it to search the internet for some information, it will not be able to do so.

This is where tools and agents come into play. Tools and agents enable LLMs to retrieve information from various external sources such as the internet, Wikipedia, YouTube, or virtually any Python method implemented as a tool in LangChain.

This article will show you how to enhance the functionalities of your RAG systems using tools and agents in the Python LangChain framework.

So, let's begin without an ado.

The following script installs the required libraries, including the Python LangChain framework and its associated modules and the OpenAI client.

!pip install -U langchain
!pip install langchain-core
!pip install langchainhub
!pip install -qU langchain-openai
!pip install pypdf
!pip install faiss-cpu
!pip install --upgrade --quiet  wikipedia
Requirement already satisfied: langchain in c:\us

The script below imports the required libraries into your Python application.

from langchain.tools import WikipediaQueryRun
from langchain_community.utilities import WikipediaAPIWrapper
from langchain.tools.retriever import create_retriever_tool
from langchain.agents import AgentExecutor, create_tool_calling_agent
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.chat_message_histories import ChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory
from langchain_core.tools import tool
from langchain import hub
import os

from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS

To enhance RAG using tools and agents in LangChain, follow the following steps.

1. Import or create tools you want to use with RAG.
2. Create a retrieval tool for RAG.
3. Add retrieval and other tools to an agent.
4. Create an agent executor that invokes agents' calls.

The benefit of agents over chains is that agents decide at runtime which tool to use to answer user queries.

This article will enhance the RAG model's performance using the Wikipedia tool. We will create a LangChain agent with a RAG tool capable of answering questions from a document containing information about the British parliamentary system. We will incorporate the Wikipedia tool into the agent to enhance its functionality.

If a user asks a question about the British parliament, the agent will call the RAG tool to answer it. In case of any other query, the agent will use the Wikipedia tool to search for answers on Wikipedia.

Let's implement this model step by step.

The following script imports the built-in Wikipedia tool from the LangChain module. To retrieve Wikipedia pages, pass your query to the run() method.

wikipedia_tool = WikipediaQueryRun(api_wrapper=WikipediaAPIWrapper())
response = wikipedia_tool.run("What are large language models?")
print(response)

Output:

To add the above tool to an agent, you must define a function using the @tool decorator. Inside the method, you simply call the run() method as you previously did and return the response.

@tool
def WikipediaSearch(search_term: str):

    """
    Use this tool to search for wikipedia articles.
    If a user asks to search the internet, you can search via this wikipedia tool.
    """


    result = wikipedia_tool.run(search_term)
    return result

Next, we will create a retrieval tool that implements the RAG functionality.

In my article on Retrieval Augmented Generation with Claude 3.5 Sonnet, I explained how to create retriever using LangChain. The process remains the same here.

openai_api_key = os.getenv('OPENAI_API_KEY')

loader = PyPDFLoader("https://web.archive.org/web/20170809122528id_/http://global-settlement.org/pub/The%20English%20Constitution%20-%20by%20Walter%20Bagehot.pdf")
docs = loader.load_and_split()

documents = RecursiveCharacterTextSplitter(
    chunk_size=1000, chunk_overlap=200
).split_documents(docs)

embeddings = OpenAIEmbeddings(openai_api_key = openai_api_key)
vector = FAISS.from_documents(documents, embeddings)

retriever = vector.as_retriever()

You can query the vector retriever using the invoke() method. In the output, you will see the section of documents having the highest semantic similarity with the input.

query = """
"What is the difference between the house of lords and house of commons? How members are elected for both?"
"""
retriever.invoke(query)[:3]

Output:

Next, we will create a retrieval tool that uses the vector retriever you created to answer user queries. You can create a RAG retrieval tool using the create_retriever_tool() function.

BritishParliamentSearch = create_retriever_tool(
    retriever,
    "british_parliament_search",
    "Use this tool tos earch for information about the british parliament, house of lords and house of common and any other related information.",
)

We have created a Wikipedia tool and a retriever (RAG) tool; the next step is adding these tools to an agent.

First, we will create a list containing all our tools. Next, we will define the prompt that we will use to call our agent. I used a built-in prompt from LangSmith, which you can see in the script's output below.

tools = [WikipediaSearch, BritishParliamentSearch]
# Get the prompt to use - you can modify this!
prompt = hub.pull("hwchase17/openai-functions-agent")
prompt.messages

Output:

You can use your prompt if you want.

We also need to define the LLM we will use with the agent. We will use the OpenAI GPT-4o in our script. You can use any other LLM from LangChain.

llm = ChatOpenAI(model="gpt-4o",
                 temperature=0,
                 api_key=openai_api_key,
                )

Next, we will create a tool-calling agent that generates responses using the tools, LLM, and the prompt we just defined.

Finally, to execute an agent, we need to define our agent executor, which returns the agent's response to the user when invoked via the invoke() method.

In the script below, we ask our agent a question about the British parliament.

agent = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose = True)
response = agent_executor.invoke({"input": "How many members are there in the House of Common?"})
print(response)

Output:

As you can see from the above output, the agent invoked the british_parliament_search tool to generate a response.

Let's ask another question about the President of the United States. Since this information is not available in the document that the RAG tool uses, the agent will call the WikipediaSearch tool to generate the response to this query.

response = agent_executor.invoke({"input": "Who is the current president of United States?"})

Output:

Finally, if you want only to return the response without any additional information, you can use the output key of the response as shown below:

print(response["output"])

Output:

The current President of the United States is Joe Biden. He is the 46th president and assumed office on January 20, 2021.

As a last step, I will show you how to add memory to your agents so that they remember previous conversations with the user.

We will first create an AgentExecutor object as we did previously, but this time, we will set verbose = False since we are not interested in seeing the agent's internal workings.

Next, we will create an object of the ChatMessageHistory() class to save past conversations.

Finally, we will create an object of the RunnableWithMessageHistory() class and pass to it the agent executor and message history objects. We also pass the keys for the user input and chat history.

agent_executor = AgentExecutor(agent=agent, tools=tools, verbose = False)

message_history = ChatMessageHistory()

agent_with_chat_history = RunnableWithMessageHistory(
    agent_executor,
    # This is needed because in most real world scenarios, a session id is needed
    # It isn't really used here because we are using a simple in memory ChatMessageHistory
    lambda session_id: message_history,
    input_messages_key="input",
    history_messages_key="chat_history",
)

Next, we will define a function generate_response() that accepts a user query as a function parameter and invokes the RunnableWithMessageHistory() class object. In this case, you also need to pass the session ID, which points to the past conversation. You can have multiple session IDs if you want multiple conversations.

def generate_response(query):
    response = agent_with_chat_history.invoke(
    {"input": query},
    config={"configurable": {"session_id": "<foo>"}}
    )

    return response

Let's test the generate_response() function by asking a question about the British parliament.

query = "What is the difference between the house of lords and the house of commons?"
response = generate_response(query)
print(response["output"])

Output:

Next, we will ask a question about the US president.

query = "Who is the current President of America?"
response = generate_response(query)
print(response["output"])

Output:

The current President of the United States is Joe Biden. He assumed office on January 20, 2021, and is the 46th president of the United States.

Next, we will only ask And for France but since the agent remembers the past conversation, it will figure out that the user wants to know about the current French President.

query = "And of France?"
response = generate_response(query)
print(response["output"])

Output:

The current President of France is Emmanuel Macron. He has been in office since May 14, 2017, and was re-elected for a second term in 2022.

Retrieval augmented generation allows you to answer questions using documents from a vector database. However, you may need to fetch information from external sources. This is where tools and agents come into play.

In this article, you saw how to enhance the functionalities of your RAG systems using tools and agents in LangChain. I encourage you to incorporate other tools and agents into your RAG systems to build amazing LLM products.

View Transitions are one of the most awesome features CSS has shipped in recent times. Its title is self-explanatory: transitions between views are possible with just CSS, even across pages of the same origin! What’s more interesting is its subtext, since there is no need to create complex SPA with routing just to get those eye-catching transitions between pages.

What also makes View Transitions amazing is how quickly it has gone from its first public draft back in October 2022 to shipping in browsers and even in some production contexts like Airbnb — something that doesn’t happen to every feature coming to CSS, so it shows how rightfully hyped it is.

That said, the API is still new, so it’s bound to have some edge cases or bugs being solved as they come. An interesting way to keep up with the latest developments about CSS features like View Transitions is directly from the CSS Telecom Minutes (you can subscribe to them at W3C.org).

View Transitions were the primary focus at the August 21 meeting, which had a long agenda to address. It started with a light bug in Chrome regarding the navigation descriptor, used in every cross-document view transition to opt-in to a view transition.

@view-transition {
  navigation: auto | none;
}

Currently, the specs define navigation as an enum type (a set of predefined types), but Blink takes it as a CSSOMString (any string). While this initially was passed as a bug, it’s interesting to see the conversation it sparked on the GitHub Issue:

Actually I think this is debatable, we don’t currently have at rules that use enums in that way, and usually CSSOM doesn’t try to be fully type-safe in this way. e.g. if we add new navigation types and some browsers don’t support them, this would interpret them as invalid rules rather than rules with empty navigation.

The last statement may not look exciting, but it opens the possibility of new navigation types beyond auto and none, so think about what a different type of view transition could do.

And then onto the CSSWG Minutes:

emilio: Is it useful to differentiate between missing auto or none?

noamr: Yes, very important for forward compat. If one browser adds another type that others don’t have yet, then we want to see that there’s a difference between none or invalid

emilio: But then you get auto behavior?

noamr: No, the unknown value is not read for purpose of nav. It’s a vt role without navigation descriptor and no initial value Similar to having invalid rule

So in future implementations, an invalid navigation descriptor will be ignored, but exactly how is still under debate:

ntim: How is it different from navigation none?

noamr: Auto vs invalid and then auto vs none. None would supersede auto; it has a meaning to not do a nav while invalid is a no-op.

ntim: So none cancels the nav from the prev doc?

noamr: Yes

The none has the intent to cancel any view transitions from a previous document, while an invalid or empty string will be ignored. In the end, it resolved to return an empty string if it’s missing or invalid.

RESOLVED: navigation is a CSSOMString, it returns an empty string when navigation descriptor is missing or invalid

Onto the next item on the agenda. The discussion went into the view-transition-group property and whether it should have an order of precedence. Not to confuse with the pseudo-element of the same name (::view-transition-group) the view-transition-group property was resolved to be added somewhere in the future. As of right now, the tree of pseudo-elements created by view transitions is flattened:

::view-transition
├─ ::view-transition-group(name-1)
│  └─ ::view-transition-image-pair(name-1)
│     ├─ ::view-transition-old(name-1)
│     └─ ::view-transition-new(name-1)
├─ ::view-transition-group(name-2)
│  └─ ::view-transition-image-pair(name-2)
│     ├─ ::view-transition-old(name-2)
│     └─ ::view-transition-new(name-2)
│ /* and so one... */

However, we may want to nest transition groups into each other for more complex transitions, resulting in a tree with ::view-transition-group inside others ::view-transition-group, like the following:

::view-transition
├─ ::view-transition-group(container-a)
│  ├─ ::view-transition-group(name-1)
│  └─ ::view-transition-group(name-2)
└─ ::view-transition-group(container-b)
    ├─ ::view-transition-group(name-1)
    └─ ::view-transition-group(name-2)

So the view-transition-group property was born, or to be precise, it will be at some point in timer. It might look something close to the following syntax if I’m following along correctly:

view-transition-group: normal | <ident> | nearest | contain;

• normal is contained by the root ::view-transition (current behavior).
• <ident> will be contained by an element with a matching view-transition-name
• nearest will be contained by its nearest ancestor with view-transition-name.
• contain will contain all its descendants without changing the element’s position in the tree

The values seem simple, but they can conflict with each other. Imagine the following nested structure:

A  /* view-transition-name: foo */
└─ B /* view-transition-group: contain */
   └─ C /* view-transition-group: foo */

Here, B wants to contain C, but C explicitly says it wants to be contained by A. So, which wins?

vmpstr: Regarding nesting with view-transition-group, it takes keywords or ident. Contain says that all of the view-transition descendants are nested. Ident says same thing but also element itself will nest on the thing with that ident. Question is what happens if an element has a view-transition-group with a custom ident and also has an ancestor set to contain – where do we nest this? the contain one or the one with the ident? noam and I agree that ident should probably win, seems more specific.

<khush>: +1

The conversations continued if there should be a contain keyword that wins over <ident>

emilio: Agree that this seems desirable. Is there any use case for actually enforcing the containment? Do we need a strong contain? I don’t think so?

astearns: Somewhere along the line of adding a new keyword such as contain-idents?

<fantasai>: “contain-all”

emilio: Yeah, like sth to contain everything but needs a use case

But for now, it was set for <ident> to have more specificity than contain

PROPOSED RESOLUTION: idents take precedence over contain in view-transition-group

astearns: objections or concerns or questions?

<fantasai>: just as they do for <ident> values. (which also apply containment, but only to ‘normal’ elements)

RESOLVED: idents take precedence over contain in view-transition-group

Lastly, the main course of the discussion: whether or not some properties should be captured as styles instead of as a snapshot. Right now, view transitions work by taking a snapshot of the “old” view and transitioning to the “new” page. However, not everything is baked into the snapshot; some relevant properties are saved so they can be animated more carefully.

From the spec:

However, properties like mix-blend-mode which define how the element draws when it is embedded can’t be applied to its image. Such properties are applied to the element’s corresponding ::view-transition-group() pseudo-element, which is meant to generate a box equivalent to the element.

In short, some properties that depend on the element’s container are applied to the ::view-transition-group rather than ::view-transition-image-pair(). Since, in the future, we could nest groups inside groups, how we capture those properties has a lot more nuance.

noamr: Biggest issue we want to discuss today, how we capture and display nested components but also applies to non-nested view transition elements derived from the nested conversation. When we nest groups, some CSS properties that were previously not that important to capture are now very important because otherwise it looks broken. Two groups: tree effects like opacity, mask, clip-path, filters, perspective, these apply to entire tree; borders and border-radius because once you have a hierarchy of groups, and you have overflow then the overflow affects the origin where you draw the borders and shadows these also paint after backgrounds

noamr: We see three options.

1. Change everything by default and don’t just capture snapshot but add more things that get captured as ?? instead of a flat snapshot (opacity, filter, transform, bg borders). Will change things because these styles are part of the group but have changed things before (but this is different as it changes observable computed style)
2. Add new property view-transition-style or view-transition-capture-mode. Fan of the first as it reminds me of transform-style.
3. To have this new property but give it auto value. If group contains other groups when you get the new mode so users using nesting get the new mode but can have a property to change the behavior If people want the old crossfade behavior they can always do so by regular DOM nesting

Regarding the first option about changing how all view transitions capture properties by default:

bramus: Yes, this would be breaking, but it would break in a good way. Regarding the name of the property, one of the values proposed is cross-fade, which is a value I wouldn’t recommend because authors can change the animation, e.g. to scale-up/ scale-down, etc. I would suggest a different name for the property, view-transition-capture-mode: flat | layered

Of course, changing how view transitions work is a dilemma to really think about:

noamr: There is some sentiment to 1 but I feel people need to think about this more?

astearns: Could resolve on option 1 and have blink try it out to see how much breakage there is and if its manageable then we’re good and come back to this. Would be resolving one 1 unless it’s not possible. I’d rather not define a new capture mode without a switch

…so the best course of action was to gather more data and decide:

khush: When we prototype we’ll find edge cases. We will take those back to the WG in that case. Want to get this right

noamr: It involves a lot of CSS props. Some of them are captured and not painted, while others are painted. The ones specifically would all be specified

After some more discussion, it was resolved to come back with compat data from browsers, you can read the full minutes at W3C.org. I bet there are a lot of interesting things I missed, so I encourage you to read it.

RESOLVED: Change the capture mode for all view-transitions and specify how each property is affected by this capture mode change

RESOLVED: Describe categorization of properties in the Module Interactions sections of each spec

RESOLVED: Blink will experiment and come back with changes needed if there are compat concerns

CSSWG Minutes Telecon (2024-08-21) originally published on CSS-Tricks, which is part of the DigitalOcean family. You should get the newsletter.

Even if you’re brand new to brand voice, you already know exactly what it is. I promise.

Think of a few of your favorite brands, and consider why they’re favs. The product or service probably has a lot to do with it, but that’s only part of the story — a brand’s voice or personality is also a major factor in consumer loyalty. 

Think about the overall vibe of your favorite brand — is it friendly? Authoritative? Funny? That’s brand voice at work.

A well-defined brand voice can underscore your authority, play up your playfulness, or simply bring the directness and relatability that consumers look for in brands. A poorly defined voice, or one that changes frequently, undermines your brand and alienates customers or clients.

So let’s talk about how to start from scratch by looking at the elements that form a brand’s voice, plus 10 examples to inspire you.

Table of Contents

• Brand Voice Examples
• What is a brand voice?
• Why Brand Voice is Important
• Creating a Brand Voice
• Brand Voice Template
• Crafting Your Voice

Your company’s voice should resonate with your audience and build trust with them. In the U.S. market, 90% of consumers say it’s important to trust the brands they buy or use.

Your brand voice shows your customers what to expect from your company’s content, services, and even customer service.

Why Brand Voice is Important

Brand voice is a little bit like a brand ambassador.

You’ll make certain assumptions about an unfamiliar brand if its ambassadors are clad in pink cowboy hats or in black three-piece suits. And you’ll know immediately whether you’re the target audience.

A brand’s voice is usually defined by four or so adjectives that immediately convey whether you’re a pink cowboy hat kind of brand (bubbly, playful, youthful, irreverent) or a black three-piece suit kind of brand (somber, formal, authoritative, exclusive).

Every bit of copy that your brand produces, whether it’s the About Us page on your website or the game on the back of a cereal box, should exude your brand’s distinct voice.

Put some thought into those four(ish) adjectives — we’ll show you how — because your brand voice has to translate across multiple platforms, and potentially even across countries and cultures.

It has an important internal function, too. A well-defined brand voice establishes a cohesive set of guidelines for your writers, marketers, content creators, and even graphic designers.

“Well-defined” is key here — you can throw a bunch of adjectives at the wall and hope something sticks, but without a solid explanation of what “clear, helpful, human, and kind” means, you’re in danger of muddied or inconsistent content.

HubSpot’s style guide, for instance, specifies that “we favor clarity above all. The clever and cute should never be at the expense of the clear.” It also gives multiple examples of what “clear,” “helpful,” “human,” and “kind” actually look like in our copy — a godsend for contractors and new hires.

Once you’ve nailed down your brand’s voice, you’ll find it easier to speak directly to your audience, attract new customers or users, and express your brand’s distinctiveness.

Creating a Brand Voice

Bring your customers into the conversation so they feel connected to your brand. If a potential customer feels like you‘re talking directly to them, then you’re doing brand voice right.

1. Start with your company's mission.

Your own values, and your company’s mission, are critical as you embark on your brand voice journey.

It’s how HubSpot’s social media team translated the brand voice to LinkedIn — and got 84% more engagement in just six months.

I asked Emily Kearns, HubSpot’s Senior Manager, Social Media, to tell me more.

“So much of what is good about HubSpot is the culture and how we treat each other — just the overall vibe,” she says. “And there was a huge opportunity to take that into the social space.”

HubSpot’s brand voice is clear, helpful, human, and kind, and Kearns says that the social media team used that as its foundation. “Human and authentic — that’s just table stakes,” she says.

But there are different ways to express clarity, helpfulness, humanness, and kindness. Where our official product descriptions might require a little more gravitas, our Instagram account can translate the HubSpot culture into ~vibes~.

Since it began reinterpreting HubSpot’s corporate voice on social media in 2023, our HubSpot social team has earned a 2024 Webby nomination in the category of Social, B2B.

Lauren Naturale, the social media manager at Tides, a nonprofit that advances social justice, agrees that values are foundational to your brand voice. “You cannot take a values-based approach to marketing if your company is not actually living or enacting those values in any meaningful way.”

Naturale was also the first social media manager at Merriam-Webster, where she developed the dictionary’s social media presence from practically nothing — “they would post the word of the day to all the social channels once a day” — into a must-follow.

She says that Merriam-Webster didn’t have the kind of strategy deck that a big corporation would have sunk a lot of money into. What it did have was “very well articulated, shared values around how interesting language was, how important it was, and the fact that it is always changing.”

She sums those values up: “Words and language are not cultural capital. They're not the property of the elite. You can care about words and language and also be interested in the way that language is changing.”

From those values, she built what is now a well-known brand voice (never mind the 456% increase in Twitter audience she ushered in).

2. Use your buyer persona as inspiration for your brand voice.

Your buyer persona should answer a few vital questions: Who are you trying to reach? What do they need from your brand? What can your brand offer them that no one else can?

Audience research can help you identify other types of content that are reliably appealing to your audience.

Tools like Google Analytics, or even a simple survey of your audience, can help you determine or confirm other sites that your readers frequent.

Ryan Shattuck, a digital media strategist who managed Dictionary.com’s social media for four years, tells me, “Knowing your audience is obvious, but I would take it a step further. Respect your audience.”

Dictionary.com’s buyer persona — or its target users — likely paints a picture of somebody who does the New York Times’ Connections word game as soon as the clock strikes midnight.

“I think it’s safe to assume that the people who follow a dictionary account on Instagram are also people who read books and do crossword puzzles,” as Shattuck puts it.

“And so I can make a joke about the Oxford comma. I can use a meme to share the etymology of a word.”

If your voice doesn't resonate with your audience, keep experimenting.

3. Look at your best-performing content.

If you've already been publishing content for a few months or even years, take a look at your top-performing pieces to find out what’s resonating with your audience.

How would you describe your brand voice in that content? It might be assured and authoritative, with deep topical knowledge backed up by original research. It could be playful and irreverent, using memes and pop-culture references to connect with your audience.

Make a list of adjectives that describe your voice in your top-performing pieces, and highlight the common elements. From there, you can start to make strategic decisions about which elements should be replicated across your brand.

It’s also helpful to research the content formats that perform the best in your industry and geographic location. (Pro tip: It’s probably short-form video.)

4. Make a list of do‘s and don’ts.

If you get stuck trying to define your brand voice, try defining what you don’t want it to be.

For instance, perhaps your team brainstorms the following statements:

Our brand voice is not pretentious.

Our brand voice is not too serious.

Our brand voice is not grandiose.

Our brand voice is not unfriendly.

Once you've taken a look at these statements, you can begin forming the antithesis. For example, the above list might yield a brand voice that’s down to earth, funny, informal, and humble.

5. If necessary, use a third-party agency to determine brand voice.

Forbes' BrandVoice is a media partnership program that helps brands reach and resonate with their audiences through expert consultancy and direct access to Forbes audiences.

Take a look at how Cole Haan worked with Forbes to create content related to style, arts, travel, social impact, and more. Each piece uses a unique voice to target the intended audience for that category.

If you're struggling to create a unique brand voice or you don’t know how to adapt your vision to the different areas of your business, consider using a program like BrandVoice or a third-party content marketing agency. This will help you take your brand’s game to the next level.

6. Create a communications document so all of your content is aligned.

Once you‘ve created your brand voice, you’ll want to ensure your entire company can use that voice in all marketing materials.

If your company only uses internal writers, consider creating a training course for new staff so they can learn how to write for your brand. If you work with external guest contributors, you'll want to make public-facing guidelines to ensure all your writing captures the appropriate voice.

7. Fill out a brand voice template with 3 - 5 core voice characteristics.

Use a table to formalize your process. Write down three to five core characteristics you‘ve determined are important for your brand’s voice and how your writers can use these traits in their writing.

This step is important for translating ideas into action — how can your writers create a “humble, authentic voice” in their writing?

Give some examples or tactical advice to make it easy for your brand voice to come through in all of your content, regardless of byline.

To explore what a template could look like in practice, take a look at the brand voice template below.

Top Tips from the Pros

Although social media is just one component of a brand’s voice, it’s often the most public and the most prolific. So I asked the social media pros I talked to for this article for their top tips on crafting a brand voice.

1. Be human.

Kearns says to ask yourself, “Would a real person say this? Is there something in here that is relatable, and that someone can connect to?”

“It’s not a dictionary sitting at a computer,” Shattuck tells me. “It’s a real person.”

Image Source.

2. Respect your audience.

It bears repeating: Don’t just know your audience. Respect them.

3. Reflect your brand’s product and culture.

You won’t win authenticity points if you’re trying to mimic another brand’s culture. Conversely, if you have a great company culture — channel it and celebrate it in your social accounts.

4. Be culturally relevant, but not at the expense of your brand identity.

This doesn’t mean you should meme-ify everything — but it does mean that memes are fair game if you stay on-brand.

Shattuck said that at Dictionary.com, he always asked himself, “Is this post educational? Is it entertaining?” If he couldn’t answer “yes” to both,, he knew the post wouldn’t do well because it wasn’t adding any value.

Brand Voice Examples

Before you start crafting your unique voice, turn to role models who have perfected their tone. Here are 10 examples to get you started.

You can see other distinct brand voices in the video below.

1. HubSpot

A year ago, you’d be more likely to find a product description on HubSpot’s social media than a meme about brat summer.

But then the social team began experimenting with a more Gen Z and millennial tone of voice.

It’s still a work in progress, Kearns tells me, and every month the team takes a close look at what performs well and what doesn’t. “We’re figuring out how we talk about the HubSpot product in a way that is interesting and adds value and is culturally relevant.”

Cultural relevance and timeliness are major considerations for the social team. Kearns says she’s always asking herself how they can connect the HubSpot product to “something hyper relevant, or something that managers are going through right now.”

“If we just talk about our product in a vacuum, even with our fun brand voice layered on top of it, it might fall flat.”

Kearns says that although your brand voice should be identifiable and consistent, “it should have a little bit of flexibility” so you can adapt it to different platforms.

2. Duolingo

Duo the owl is the face that launched a thousand memes.

Image Source

The feathery embodiment of the Duolingo brand voice, Duo is “expressive, playful, embracing, and worldly.” That’s according to Duolingo’s brand guide, which also notes that Duo is both “persistent” and “slightly awkward.”

Duolingo’s defined brand voice includes a “brand personality” section that describes who Duolingo would be as a celebrity (Trevor Noah), a vehicle (a Vespa), and a song (Queen’s “Don’t Stop Me Now”).

Duolingo’s Senior Global Social Media Manager, Zaria Parvez, told Contagious in a 2023 interview, “Dream big, but iterate small.”

If you’ve spent any time on the clock app, you’re familiar with Duo’s occasionally unhinged antics — which all started with Parvez asking to take over Duolingo’s then-dormant TikTok account.

3. Title Nine

A woman-owned and women-focused athleticwear company, Title Nine combines a friendly “aww shucks” vibe with a triumphant fist pump.

Freelance copywriter Robyn Gunn writes on her website that T9 brought her in to write copy that “reinforce[s] the brand's badass, ballsy DNA that differentiates it from ‘softer’ competitors in the category.”

Title Nine’s “Who We Are” page encapsulates this voice perfectly: It’s written in clear, simple language that underscores the brand’s love of the outdoors and its enduring support of women.

Image Source

This graphic from its online store brings out a more playful side of Title Nine’s brand voice, evident in the bright colors and patterns, the casual typeface that “Trail Shop” uses, and the invitation to “track in some dirt.”

Image Source

Title Nine doesn’t have a publicly accessible brand guide, but I’d describe its voice as friendly, powerful, playful, and direct.

4. Who Gives a Crap

True story: A customer service rep at Capital One once had to read me a list of recent credit card charges so I could confirm whether they were mine or a fraud.

Poor dude was clearly mortified at having to read “Who Gives a Crap” out loud, saying, “This is the company name, I am just reading this off a list, it is not me saying this.”

So he’s maybe not WGaC’s target audience, which is considerably more relaxed on the topic of toilet paper.

WGaC’s “About Us” page tells a tale of toilet jokes and changing the world. Successfully combining something so ridiculous with a very real and very serious global problem is no easy task, but the ability to walk that line nicely sums up the brand’s voice.

Image Source

“Making a difference in the world” can be a hard value to channel in a brand voice, since the brand (and the people behind it) have to demonstrably live up to the promise of effecting change.

Who Gives a Crap gives a lot of specific details that indicate that lack of access to a toilet is an issue that the founders genuinely care about. The product descriptions do the same. Take this one for a special poetry edition TP (I am not making this up):

Image Source

“Create an ode in the commode” is pretty hard to beat for terrible poetry. The product description ends with, “And since we donate 50% of profits, you’re not just building ballads, you’re doing good, too!” — a reminder of the brand’s promise in a goofy, casual tone.

WGaC’s brand voice might be described as cheeky (pun absolutely intended), lighthearted but rooted in a cause that’s deeply serious, informal, and conversational.

5. Poppi

Poppi soda blares its voice from the moment you land on its eye-searing bright pink and yellow website. Known for having a Gen Z-friendly voice, DrinkPoppi.com looks more like a neon Instagram feed than a website for flavored sparkling water.

Its “About Us” page brags about “new besties” like Billie Eilish and Post Malone, and even its newsletter sign-up says, “Let’s be friends.”

Image Source

The creative agency responsible for Poppi’s branding describes “the world of Poppi” as “quirky, nostalgic, and vibrant.” I’d add to that “informal” or “casual.”

6. Spotify

Whether you‘re watching a TV ad, driving past a billboard, or scrolling Spotify’s social accounts, you'll see a consistent voice. The brand’s tone is consistently funny, edgy, direct, and concise.

For instance, take a look at this video, which is part of a Spotify advertisement campaign from 2019, “Let the Song Play.”

As you can see, Spotify doesn‘t take itself too seriously. The ad makes fun of people who get so emotionally invested in a song that they won’t resume their plans until the song ends.

You‘ll see a similar brand voice play out on Spotify’s social channels. On its Twitter account, for instance, the brand often posts tweets related to new music in a casual, friendly manner.

Image Source

If Spotify‘s brand were a person, she would be witty, sarcastic, and up-to-date on today’s pop culture references. You‘ll see that personality play out across all of Spotify’s communication channels.

7. Mailchimp

When exploring Mailchimp's brand voice, turn to the company’s Content Style Guide.

In the Style Guide, Mailchimp writes, “We want to educate people without patronizing or confusing them. Using offbeat humor and a conversational voice, we play with language to bring joy to their work.… We don't take ourselves too seriously.”

Image Source

Even in the Style Guide, you can hear Mailchimp's brand voice shine through. The company consistently achieves a conversational, direct, playful voice in all its content.

For instance, in this blog post, the brand writes about various “highly unscientific personas”, including the fainting goat. The email service provider describes this persona by saying, "when startled, its muscles stiffen up and it falls right over.”

They then link out to this hilarious video.

As you can see from this example, you can evoke brand voice in subtle yet effective ways. If the blogger had instead written, “If a goat is scared, it becomes nervous. The animal's muscles contract and it faints as a result”, the writer would've evoked a voice more aligned with a scientific journal than Mailchimp.

Image Source

8. Fenty Beauty

The About Us page for Rihanna's beauty company reads, "Before she was BadGalRiRi: music, fashion and beauty icon, Robyn Rihanna Fenty was a little girl in Barbados transfixed by her mother’s lipstick.

The first time she experienced makeup for herself, she never looked back. Makeup became her weapon of choice for self-expression."

It‘s clear, even just through this short snippet, that Fenty Beauty’s voice is bold, direct, and poetic. Language like “transfixed by her mother's lipstick” and “her weapon of choice for self-expression” reinforce this voice. However, the tone is also undeniably casual — the way you might talk to your best friend.

Image Source

You'll see this voice play out across all Fenty social channels, including this YouTube video description:

Image Source

The first statement, “The blur is REAL!” — along with phrases like “No-makeup makeup look”, and the shortening of the word “combination” — all evoke a sense of friendliness.

The brand voice matches its target audience perfectly: youthful millennials and Gen-Zers who care about makeup as an opportunity for authentic expression.

9. Clare Paint

Clare, an online paint site, has created a mature, spirited, and cheerful brand voice to evoke a breezy, girl-next-door feel to their branded content.

For instance, consider the title of one of their recent blog posts, “6 Stylish Rooms on Instagram That Make a Strong Case For Pink Walls.”

The post uses phrases like “millennial pink”, “pink walls have obvious staying power”, and “designers and DIY enthusiasts alike have embraced the playful shade with open arms.”

The brand’s language is friendly, chic, and professional, relating to its readers while simultaneously demonstrating the brand's home decor expertise.

Image Source

This voice is clear across channels. Take a look at this Instagram post, for instance.

Image Source

“When baby's first bedroom is on your grown-up vision board” makes the brand feel like a good-natured older (and more fashionable) sister. The reference to the COO‘s baby boy is another opportunity to make authentic connections with Clare’s followers.

10. Skittles

Skittles often posts hilarious social media posts that strip away any promotional, phony language so you're left with something much more real.

Take this tweet, which reads: “Vote Skittles for Best Brand on Twitter so we can keep our jobs!”

Image Source

The brand voice, which is clever and original, does a good job of making prospects and customers feel like they‘re chatting with a mischievous employee behind-the-scenes. The “I can’t believe they just posted that” factor keeps the content fresh and exciting.

Plus, the brand does a good job making pop culture references, like this Mean Girls reference, to highlight the brand's youthfulness.

Image Source

Skittles’ use of absurdity and humor plays into their iconic commercials. In one 2022 ad, the company pokes fun at targeted ads.

While two people watch a youtube video, they comment that their ads are so targeted that it feels as if Skittles is listening in on their conversation. Then, a man with a boom mike drops through the floor.

Skittles expertly keeps the same tone across media, showing their brand’s commitment to their voice.

Brand Voice Template

Image Source

Looking to make a template for your own brand voice? HubSpot is here to help! You can fill out this blank Google Sheet template with your own brand voice characteristics.

Fill out the remaining cells, and send them along to your team.

It‘s important to note, you’ll be prompted to make a Google Drive copy of the template, which isn't possible without a Google account.

Crafting Your Voice

And there you have it! You're well on your way toward building a strong, compelling brand voice for your own business.

Logo, color palette, and font are all important aspects of branding. But beyond that, a good brand starts with good content. And good content can’t exist without a strong voice.

Editor's Note: This post was originally published in April 2021 and has been updated for comprehensiveness.

Influencer marketing is one of the most powerful and most effective forms of digital marketing available today.

With a market value of over $21 billion dollars, over 3X the value since 2019, influencer marketing continues to grow as it provides a level playing field for both big and small brands to compete for the user’s attention.

According to a survey by Matter Communications, 81% of people claimed they’ve made purchases based on recommendations from influencers, friends, and family.

Unlike other traditional forms of marketing, influencer marketing evolves with new trends, strategies, and concepts. Each year, you’ll see something new and different that brings better results for your campaigns.

In this post, we take a deep dive into some of these newest influencer marketing trends to see how you can leverage them to get better returns on investment (ROI). These methods might not be as effective next year so it’s better to take advantage of them right now!

1. Rise of AI Influencers

It’s no secret that most influencers on social media platforms now utilize some form of AI in their content creation process to make everything look perfect, whether it’s writing better captions, generating AI images, or even enhancing their everyday photos with a little bit of AI editing.

However, real influencers will have to do much more than that to stay relevant as they now have new competition from an entirely new group of influencers that are even more seemingly perfect.

Miquela is just one of many AI influencers currently taking over social media platforms. With more than 2.5 million followers, this AI influencer is not even the biggest account of its kind.

While it’s a new concept, most of these AI influencer accounts are already creating sponsored content and brands do not hesitate to get their products reviewed by an AI personality.

The same trend can be seen across all social media platforms from VTubers on YouTube to AI models on Instagram and TikTok. Facebook is also working on a new AI personality model that allows creators to make AI representations of themselves.

At this rapid rate of advancements in AI, it’s difficult to predict how big of an impact it will have on social media.

2. TikTok Domination

60-second videos are still the “king” of social media marketing and TikTok is still in the lead on that front. Especially when it comes to targeting Gen-Z audiences, TikTok is the go-to platform for marketing campaigns.

(Source: Statista)

Even though TikTok ranks at the 5th position in terms of the number of monthly active users, it is the most popular platform among brands for influencer marketing. 66% of brands use TikTok for their influencer marketing campaigns, with only 47% using Instagram, 33% YouTube, and 28% Facebook.

Even though Instagram, YouTube, and Facebook have also implemented their own versions of short-form content formats, TikTok still dominates the Gen-Z market with its trend-making content creation system.

3. Growth of Nano-Influencers

Gone are the days of looking up to celebrities for fashion and lifestyle advice. Now, younger audiences, especially Gen-Zers, turn to smaller influencers for fashion advice.

According to another study by Matter Communications, 60% to 70% of consumers prefer advice from small influencers who are experts in their field while only 17% to 22% listen to popular celebrities.

(Source: TikTok)

In fact, the smaller the follower count the better. Engagement rates were much higher for influencers with 1,000 to 10,000 followers (nano-influencers) than influencers with 10,000 to 50,000 followers (micro-influencers).

The biggest reason for this is authenticity. Smaller influencers put more effort and energy into creating more original content and having closer relationships with their audience.

4. Niche Trends vs Global Trends

Going viral on a social media platform is still a shot in the dark. With changing algorithms and audience interests, no one can predict what the next meme or trending hashtag could be. While big brands are still making bets on these global trends, smaller brands and businesses are now slowly turning towards niche trends.

(Source: Google Trends)

Local and niche trends are much easier to target than taking a shot at a trending global hashtag or an event. Small brands have a much higher chance of going viral on a small scale with niche trends, making it easier to even target specific local audiences.

Working with small, local influencers is the key to success in niche trends. Finding micro and nano influencers who specialize in a niche market will make it much easier to target your ideal audience and create trendy content that goes viral.

5. Long-term Campaigns

Influencer marketing is a marathon, not a sprint. You can’t convince your target audience to buy your product over a more established, competing product with just a single TikTok video. It takes long-term campaigns and partnerships to achieve that goal.

This is why many businesses now set aside marketing budgets specifically for content and influencer marketing. According to a HubSpot report, 50% of marketers plan on increasing their content marketing budgets this year.

(Source: YouTube)

As a result, most brands are now establishing long-term partnerships with influencers. Some even go as far as making ambassador-like partnerships that allows influencers to promote products exclusively from a single brand.

It’s an effective way to win over an audience rather than one-time sponsorships from an influencer who promotes products from different brands without consistency.

6. Livestream Shopping & User-Generated Content (UGC)

Leveraging user-generated content (UGC) in social media marketing is nothing new. It’s been an effective strategy for many years. However, a new trend in UGC offers a more effective, multi-layered approach that brings much higher benefits.

It involves partnering with influencers to encourage their audience to create content about a product. Usually, a giveaway or some sort of a prize is necessary for this process, sometimes even asking the audience to involve their friends as a part of the process.

This type of marketing greatly benefits the brand as it generates buzz around the product and the brand on multiple levels.

Livestream shopping plays a huge role in this process as well. Platforms like Whatnot are allowing content creators and influencers to easily create shopping experiences and involve their audiences in the process.

7. Influencer Seeding

Sending influencers free product samples, also known as influencer seeding, is still an effective strategy used in social media marketing. While it’s not as engaging as sponsored partnerships, this strategy allows much smaller businesses without big marketing budgets to create a presence for their brand.

(Source: TikTok)

A better way to approach this method is to create personalized packages that show the influencers how much you appreciate them and what they are doing. Including several additional free product samples for a giveaway will also get their attention.

Conclusion

As you can see, there are many different opportunities and ways to approach influencer marketing. Big or small, every brand can use these influencer trends to create awareness and reach new audiences. Be on the lookout for the latest trends and leverage them to beat the competition.

Experimentation is central to making evidence-based decisions, and this is where A/B testing has always shined.

But with the advent of AI, we now have tools for AI A/B testing, making experimentation smarter, faster, and infinitely more manageable.

AI A/B testing gets you real-time reports and lets you test multiple hypotheses in a few clicks. To explore the magic that AI brings to A/B testing, I spoke with CRO experts who shared their unique insights.

On top of that, I’ll also take you through the benefits, limitations, and best practices for integrating AI into your A/B testing process.

In this article:

• Why use AI for A/B testing?
• Benefits of AI in A/B Testing
• Limitations of AI in A/B Testing
• How to Use AI for A/B Testing

Why use AI for A/B testing?

A/B testing is a research method used to analyze landing pages, user interfaces, or other marketing prototypes to determine the best version before full rollout.

You split your audience into two groups or more. One sees the control (A; original version), while the other interacts with the variant (B; modified version). Tracking interactions, analyzing results, and refining content follows.

With AI, you automate much of this heavy lifting. You get clear, actionable insights without the usual headaches because AI takes the guesswork out of the following:

• Testing idea development. AI systems, particularly those using machine learning like ChatGPT, can sift through massive datasets. They can help generate fresh test ideas and refine suggestions as you amass more data. Need inspiration? I like this Advertising A/B Testing ChatGPT prompts created by advertising agency Anything is Possible Media Ltd.

Image Source

• Data modeling and analysis. Quality data is the foundation for solid and reliable A/B tests. AI helps by cleaning data, i.e., removing errors, duplicates, and inconsistencies that could skew test results.

• Test customization. Say you have a mix of local and foreign visitors on your site. A 50/50 split may only attract local traffic since perks requiring in-store visits won’t appeal to international shoppers. AI ensures this testing only reaches locals.

• Testing process. AI systems like VWO set up experiments, track user interactions in real-time, analyze performance metrics, and offer suggestions for improvement. This automation reduces manual effort and speeds up testing cycles.

• Variant generation. Instead of manually creating each test version, AI generates new variants based on your criteria. It tests multiple ideas at once and prioritizes the most promising ones.

Artificial intelligence can help you sidestep the usual pitfalls of human-led A/B testing. Here’s how AI and traditional methods stack up against each other.

With AI handling everything from setup to analysis, you can ditch the old-school grind for clearer, faster insights. Let’s explore how these efficiencies benefit your A/B testing strategy and set you up for success.

Benefits of AI in A/B Testing

AI streamlines your workflow and generates more accurate insights faster. Here are the top benefits that make AI indispensable for A/B testing.

Faster, Broader Data Reach

Humans take days or even weeks to gather and analyze data. Meanwhile, AI processes heaps of variables — think hundreds of web pages or app feature versions — at lightning speed.

Jon MacDonald, CEO of The Good, has reaped the benefits of this well-oiled efficiency:

“Since we build rapid prototypes quite often, using AI has helped us code A/B tests faster and without bugs. We’re able to produce rapid prototypes quickly, increasing our testing volume and rapidly validating hypotheses.”

AI distinguishes subtle correlations within large datasets, helping you prioritize and evaluate the right variants. Thus, you get results faster and make smarter decisions without getting bogged down by lengthy analysis.

Improved Accuracy

Manual error and cognitive biases can skew the results and interpretation of A/B tests. This study on advertising A/B testing demonstrates how AI improves accuracy in these four dimensions:

1. Targeting. Machine learning lets you create detailed audience segments. Some AI tools even allow for real-time, targeted adjustments based on live data.

2. Personalization. Using Recommendation System and Virtual Assistant technology, AI tailors content to individual preferences. Each A/B test variation only shows up for users with similar interests.

3. Content creation. Generative AI and Natural Language Processing (NLP) enhance ad content quality and diversity. You can leverage it to generate consistent, high-quality ad variations.

4. Ad optimization. Deep Learning and Reinforcement Learning adjust advertising strategies dynamically. It optimizes factors like ad placement, timing, and frequency based on live performance data.

AI improves accuracy at every stage of A/B testing. It fine-tunes your test parameters, ensures optimal testing for all variants, and provides deeper insights into user interactions.

Predictive Capabilities

AI doesn’t stop at analyzing past data. It also predicts future trends to forecast how users respond to changes and make proactive adjustments.

Advanced tools such as Kameleoon use historical data and predictive analytics to anticipate visitor behavior. Kameleoon achieves this with its Kameleoon Conversion Score (KCS™).

If KCS™ predicts visitors browsing high-end products are more likely to convert with Layout A, it ensures they see this layout. Those who are more interested in budget-friendly options may often encounter Layout B.

Your A/B tests aren’t static with AI. You’re not waiting to tweak your tests for next time. Instead, you’re optimizing and delivering the best possible experience instantaneously.

Personalization

Intelligent systems track each visitor’s browsing patterns, purchase history, and preferences. AI leverages this data to tailor variations specifically for different user segments, making A/B tests more relevant and accurate.

Ashley Furniture achieved these outcomes with AB Tasty’s AI-powered platform. According to Matt Sparks, the eCommerce Optimization Manager, their UX teams used it to better understand customer experiences, solve problems, and design new functionalities.

AB Tasty helped cut out Ashley Furniture’s redundant checkout procedures. They tested a variation, prompting shoppers to enter their delivery information right after logging in. This tweak increased conversion rates by 15% and cut bounce rates by 4%.

AI-optimized test results drive tangible benefits — no doubt — but they’re not a cure-all. There are inherent limitations to consider, and we’ll go over them in the next section.

Limitations of AI in A/B Testing

AI can’t solve every problem or guarantee 100% perfect results. Recognizing the human-focused aspects it doesn’t cover allows you to be more prudent in your testing and avoid over-reliance.

Complexity

AI setup involves using advanced algorithms, specialized software, and a skilled technical team. This complexity is challenging for smaller organizations or those without a dedicated data science team.

Start with no-code platforms like Userpilot and VWO if coding isn’t your strong suit. Or, opt for out-of-the-box solutions with multi-channel support like ​​HubSpot if you test across various platforms.

Managing and optimizing A/B tests is much easier with the right tool. So, take the time to assess your needs and select a solution that aligns with your goals.

Privacy and Safety

A 2024 report by Deep Instinct shows that 97% of organizations worry they’ll suffer from AI-generated zero-day attacks.

A zero-day attack exploits a software or hardware vulnerability developers don’t yet know about, leaving no immediate fix.

If such attacks compromise your testing tools, hackers may gain unauthorized access to sensitive data. They may manipulate test results to mislead your strategy or, worse, steal users’ personal information.

Set up real-time monitoring to catch suspicious activities and implement a data breach response plan. Don’t forget to train your team on data security best practices to keep everyone vigilant.

Misinformation and Ethical Concerns

AI has no empathy and intuitive understanding. It can tell you what’s happening, but it can’t always explain why.

Tracy Laranjo, a CRO Strategist quoted in this Convert piece on AI, mentioned that AI doesn't comprehend emotions and context as humans do. She advised:

“The key is to use AI responsibly; I use it to process data more efficiently, automate repetitive tasks, and be a more concise communicator. I embrace it for the doing aspects of my job but never for the thinking aspects.”

Pro tip: Combine A/B testing with other data analysis methods or run multiple tests to gather more insights if need be. However, continue applying sound judgment when interpreting results and making decisions.

How to Use AI for A/B Testing

Below are seven ways AI can transform your A/B testing efforts.

1. Real-Time Data Analysis to Enhance Decision-Making

AI-powered A/B testing platforms can process extensive real-time data insights. They identify complex trends, patterns, and other variables, facilitating more precise tests.

One test design that exemplifies AI real-time analysis is Multi-Armed Bandit (MAB) algorithms. It allocates traffic to better-performing variations up-to-the-minute—think ad placement optimization and content recommendation.

MAB allocates ad impressions in real-time, prioritizing ads that show better performance as user data accumulates. It can also adjust content recommendations based on recent viewer interactions.

Amma, a pregnancy tracker app, used nGrow’s MAB algorithm to reduce user turnover. MAB automated and optimized push notifications in real-time, increasing retention by 12% across iOS and Android users.

The team also gained a better understanding of their user base. They can now better plan for new regions and optimize user engagement.

2. Predictive Analytics to Boost Accuracy

AI predictions prevent you from having misguided hypotheses and testing ineffective variants.

Alun Lucas, Zuko’s analytics managing director, told me how he does it. He used AI tools like ChatGPT to analyze Zuko’s form analytics data and identify the answers to the following questions:

• What are my most problematic form fields?
• How has the data changed since the last period?
• What ideas could we explore to improve the user experience and reduce abandonment in the identified problem fields?

Predictive analytics identify issues in your data forms or user flows before they become major headaches.

3. Personalized Testing to Create Tailored Experiences

AI lets you break down your audience into different segments based on behavior, demographics, and preferences.

For instance, if you plan to recommend fashion products, you can tailor your A/B tests to different customer segments. Think the patrons, bargain hunters, and eco-conscious shoppers.

Ellie Hughes, consulting head at Eclipse Group, found this approach to be valuable for validating prototypes before implementing them on a larger scale.

She tested different algorithms like personalized search ranking and photo-based recommendations. The outcome? It enhanced her clients’ experience and made it a compelling case for further AI investment.

As Hughes notes, “The value wasn’t in the production of an algorithm as an output. It was about the clever framing of an experiment to prove the monetary value of using AI within experiments.”

4. Multivariate Testing to Reveal Useful Insights

A/B testing can scale from only A and B to a full A-Z spectrum of possibilities. In her talk, Ellie Hughes debunked the myth that A/B testing is limited to comparing two versions, saying:

“A/B testing can involve multiple variants and more complex experimental designs, such as multivariate testing [...] to optimize various elements simultaneously.”

Here are some real-world instances where you can implement multivariate testing.

• Ecommerce website. Test different combinations of headlines, images, and buttons on product pages to increase conversions.
• Email marketing campaign. Experiment with subject lines, images, and call-to-action buttons to boost open and click-through rates.
• Subscription service. Try different pricing plans, promotional offers, and trial lengths to attract new customers.

Simultaneous evaluation of multiple variables offers a more nuanced approach to experimentation. It provides richer insights and better overall results than basic A/B testing.

5. Anomaly Detection to Maintain Integrity

Ever had A/B test results that seemed too good (or bad) to be true?

That happens.

Good thing is, AI tools can monitor test data 24/7 and flag any unexpected deviations from the norm. Whether it is a system glitch or a shift in user behavior, AI tools can help you diagnose these issues.

Valentin Radu, Omniconvert CEO, explained how his team used AI to understand what frustrated his clients’ customers.

They monitored NPS survey responses pre- and post-delivery. The analysis allowed his team to run more effective tests and make targeted improvements.

Radu said, “You can’t come up with strong hypotheses for your A/B tests without blending qualitative data in your insights. So, we are already using NLP to crunch the data and identify the main issues by analyzing customer feedback or survey responses.”

To formulate stronger hypotheses, cross-check quantitative data with qualitative insights. It’ll help ensure the observed anomalies aren’t due to data errors or temporary glitches.

6. Improve Search Engine Results Ranking

AI A/B testing allows for precise measurement of how different factors (e.g., algorithm changes, user interface elements, or content) impact search engine results.

Ronny Kohavi, a world-leading AI A/B testing expert, has extensively mastered online controlled experiments. His work shows how AI and machine learning have been employed for years to fine-tune search results rankings.

These rankings span major websites like Airbnb, Amazon, Facebook, and Netflix.

He informed me that Airbnb’s relevance team delivered over 6% improvements in booking conversions. That’s after 20 successful product changes out of over 250 A/B test ideas.

Kahavi says that “it's important to notice not only the positive increase to conversion or revenue but also the fact that 230 out of 250 ideas — that is, 92% — failed to deliver on ideas we thought would be useful and implemented them.”

7. Continuous Optimization to Refine A/B Tests

You tested a bold red “Buy Now” button and saw a high conversion rate last year.

Now, you notice its performance slipping. Without continuous optimization, you might not discover that users now respond better to interactive elements like hover effects or animated buttons.

Of course, these are all hypothetical scenarios, but the bottom line is clear: Continuous AI monitoring can keep your A/B tests relevant and effective.

As described in this case study, [24]7.ai continuously refined its customer service strategies through A/B testing. They tested AI-driven chat solution versions to see which improved customer interactions and resolved inquiries better.

The results? A 35% containment rate, an 8.9% bot conversion rate, and over $1.3 million saved from enhanced efficiency.

A/B test results plateau or even decline as user preferences evolve. Adjust your test parameters to keep up with changing trends and drive ongoing improvements.

Make your A/B testing smarter with AI.

AI is here. Companies and industry experts who’ve embraced AI-driven A/B testing have found it nothing short of transformative.

To get started with AI-focused A/B testing, I highly recommend checking out HubSpot’s complete A/B testing kit. It offers a comprehensive checklist to help you run the perfect split test, from initial planning to final analysis.

Now, experience the future of testing.