TL;DR
- Brands are measuring AI visibility without tracking if the perception of their brand is even accurate
- We recommend building a brand canon, which tracks facts, differentiators, brand claims, what your brand wants to be known for, and what you don’t
- Once you’ve created your canon, you can use it to compare what you want AI to say about your brand versus what it actually says (and fix the output by tracing claims to their source)
- Learn how to build your own brand canon, how to measure the accuracy of claims made by AI, and how to prioritize fixing falsehoods and inaccuracies
What Happens if AI Shares Inaccurate Info About Your Brand?
A potential customer asks AI what your company does, and the answer comes back about 70% right (per Seer's analysis of 28,123 responses across 6 AI models). But nearly 30% of the information is wrong or outdated.
That includes a product name you retired two years ago, a differentiator that belongs to your closest competitor, and a positioning line you replaced at your last rebrand.
Nobody on your team saw the answer or knows where the old information came from, and yet it just went to a prospective buyer.
So how do you fix it? None of your strategy or process documents help you answer the question "is what LLMs are saying about our brand correct?"
What feels like one bad answer is often a measurable pattern. The good news is that patterns can be diagnosed, sourced, and fixed. We’ve found brands can ask the same question across every LLM platform your buyers use, on a schedule, and get a stable read on how each one describes you.
First, you just need something to measure against. That's what we’re calling a brand canon.
A Brand Canon Is Not Your Brand Guidelines
Everyone hears "brand canon" and reaches for something they already have: brand guidelines, or a style guide, or a mission statement.
All of those serve useful but different purposes from a brand canon:
- Brand guidelines govern how your brand dresses: which colors and fonts, how much clear space around the logo, when to use the wordmark
- Style guides govern the language your brand uses: whether you use Oxford commas, capitalize "cloud", or include a dash in “e-commerce”
- History, mission, and values statements share the essence of what your brand is: how humans should think about your brand, and what judgment calls they should make
A brand canon governs what's true about your brand and what you want to be known for. It’s particularly handy in the age of LLMs and AI, because you’re putting into writing what facts, products/services, and statements you want to be associated with your brand.
But inside most companies, nobody owns the brand canon. It’s never been anybody's job to maintain a list of what is true about the brand or what your brand wants to be known for…in a form you could grade an answer against.
LLMs Position Your Brand Based on What’s Available
My colleague Alisa made the case for a brand canon in a recent blog: fix how LLMs see your brand before you chase category rankings. Build the canon, establish accuracy as your leading indicator, and investigate inaccuracies like a journalist. That post is the foundation, and the attribute list in it is where brands should start: company facts, leadership, locations, services, what you don't offer, recent rebrands.
Our first version of a canon did exactly that. We tracked and dated every citation pulled from an official source. Our documentation was good, better than what most brands have. Then we realized that version was too reactive, because it only canonized claims that were already in writing.
Your headquarters is verifiable. Your CEO's name is verifiable. When a model gets those wrong, you can point at the correct page and prove it.
But most brands don’t publicly share the position you’re trying to occupy in the market:
- The customer you want to be the obvious choice for
- The problem you want to be the first name associated with
- The reason a buyer should pick you over competitors
And if it isn't in your literature, the models have nothing to learn it from. Instead, they’ll just assign you a position.
What to Put in Your Brand Canon
When building your brand canon, we recommend focusing on five areas: facts, what you want to be known for, how your brand should be described, differentiators, and falsehoods.
Mark every entry as sourced (the entry came from a documentable source) or declared (the entry is based on your internal positioning and brand claims).
Each entry type has a different failure point: a sourced claim that fails means the model is reading the wrong page, while a declared claim that fails might mean nobody outside your company has ever heard it.
The highest-value input in the whole document is the things that are false, but that’s also the least intuitive section to build. It requires writing down things you'd never put in marketing material.
Luckily, sales already has most (if not all) of the answers: ask your team what buyers get wrong about you and what your salespeople say to correct those inaccuracies.
How to Decide if a Claim Belongs in Your Canon
Here’s the test for whether something belongs in your canon: ask yourself if you can read an AI response and rule it as accurate or inaccurate without arguing about interpretation. If not, rewrite it or cut it.
"We're the leader in customer experience": This can’t be considered a claim because nothing can contradict it, which means nothing can validate it either.
"Acme Platform supports on-premises deployment natively and is not cloud-only": This is a claim. A response either matches, or it doesn't.
You can run this test on your own messaging in about ten minutes: take your boilerplate, break it into individual statements, and mark each one as testable or untestable. Most teams find that the majority of what they say about themselves cannot be graded.
The Challenge with Declared Claims
Declared claims are harder to format than factual ones, and they break the falsifiability test if you write them the way most positioning gets written.
"We want to be known as the enterprise-grade choice for regulated industries": This is a mood; nothing can contradict it.
Turn it into a statement about what a correct answer looks like:
"In responses about compliance-heavy deployments, the brand should appear in the consideration set and be described as purpose-built for regulated environments, not as a general-purpose tool."
Declared claims also need an owner and a date, because there's no external source to cite and someone has to be accountable for the assertion. A declared claim that keeps failing isn't always a model problem. Sometimes it's telling you the position exists nowhere outside your company, and the fix is to publish that claim.
What to Do When Your Own Sources Disagree
Your canon is a record not just of what’s true, but of where your footprint contradicts itself. You’ll come across this during the first build, and how you handle it separates a real canon from a tidy one.
Consider a scenario where you have two official pages with different taglines. One is from your current positioning, the other survived a rebrand nobody finished rolling out.
Don’t just pick the one you like and move on. Record both, with sources and dates, and then flag the discrepancy as something worth tracking. You’ll likely have a preference about which version should be in circulation.
If a model repeats the old tagline, the model isn’t wrong; it just found your stale page (and now you know what to fix).
How Testing Solves the GEO Measurement Gap
Now that you know how to build your brand canon, let’s talk about how to measure and refine it.
Generative engine optimization has been missing a measurement layer that tracks not just whether you appear, but if your brand information that appears is accurate.
If you write software, you already know why this matters. A spec you can't test is a wish, and a failing test suite is only useful to tell you which failures block the release and which ones you file for the next sprint.
To solve this measurement gap, we’ve come up with our own version of a spec, a test suite, and a test run:
- The canon is the spec. It tracks what is true about your brand and who you want to be, sourced, dated, versioned.
- The claim inventory is the test suite. The spec is rewritten as assertions that can pass or fail. Each one is tiered by how much it costs you to get wrong.
- The accuracy score is the test run. It tells you which assertions failed, on which platforms, and how badly.
Not all failing tests are of equal severity. A model describing you in the wrong category is a different priority than a model naming the wrong VP of Marketing. Both are inaccuracies, but only one of them might impact revenue.
Predict the Failures Before You Start Measuring
We recommend adding one more field in your canon that changes the posture of the whole exercise: for each claim, write down what you expect models to get wrong and why.
You might predict failures like:
- Retired product names that still exist on third-party comparison pages
- Old category framing that exists in a top-cited analyst report
- Your deprecated tagline on pages you forgot to update
When you start measuring, you find out which predictions held and which didn’t. The ones that surprise you are worth investigating, because they point at sources you didn't know were shaping the answer.
Deprecated names are the failure mode to expect first. Models trained on older content will repeat what that content said. Your rebrand may have happened in your head, on your site, and in your campaigns, but that doesn’t always extend to the training corpus.
What Scoring Against a Canon Tells You
Once you run a consistent set of prompts and score each response against the claims, we suggest using four dimensions to analyze the results:
- Factual correctness: Are the stated facts true? Products, markets, leadership, scale, deployment model.
- Canon alignment: Does the response describe you the way the canon says you should be described? If every fact checks out but the model frames you inside a category you left three years ago, that's a canon alignment failure.
- Attribution quality: When a response cites sources, are they yours, third-party, or a competitor's page describing you? This is where you find out whether your own content is in the conversation.
- Hallucination presence: Are there claims about you that don’t exist anywhere? Think invented features, fabricated executives, or customers you don't have.
The output is an accuracy rate that converts "the AI says weird stuff about us sometimes" into something you can trend, compare against competitors, and hand to a CMO who is already asking.
The tiering is what makes it actionable. Once the data shows that you have a problem, weighting by what each miss costs tells you what to fix first.
What to Do When the Accuracy Rate Is Low
Trace the sourcing before you touch anything. Pull the responses that failed and look at what they cite. The answer is usually unglamorous: a five-year-old comparison page, a directory listing nobody updated, or a competitor's "alternatives to" post that frames you in their terms.
Then split the fixes, because they run on different clocks:
- Retrieval fixes work now. When a platform searches and cites live sources, you're competing for what gets retrieved. Corrected content, updated third-party pages, and pages that answer the exact question the failing prompt asks can move responses in weeks. Spend time here first, because this is where LLMs go to retrieve information.
- Corpus fixes work later. The knowledge baked into a model during training does not change the day you publish a post. What you can influence is what the next cycle ingests: the volume, consistency, and authority of accurate content across the web, including third-party sources that carry more weight than your own domain. That compounds over quarters.
Anyone promising to change training data is selling you something. What you can change is what's true about your brand's footprint on the open web, and whether you're well–positioned when the next training cycle comes around.
Once that happens, it’s time to re-score. Use the same prompts to measure your same claims against a new window.
The Limitations of Scoring Your Canon
This scenario has two limitations, and I’m calling them out now because a savvy reader will discover them anyway:
Why Everyone Skips the Canon
Most AI visibility work stops at whether you appear. We’re talking share of voice, mention rate, and citation counts.
But visibility without accuracy is being wrong at scale. Brands aren’t winning if they’re mentioned in most AI responses yet described incorrectly in half of those responses. In fact, those brands are losing faster than a brand nobody talks about, because the model is doing the describing and the buyer trusts the model.
Your canon is the difference between showing up period and showing up as yourself. You can’t measure accuracy without a ground truth. Your website is not a ground truth. Build the spec, write the tests, then find out what's failing.
You can build a canon yourself with what's in this post. Define your claims, turn them into testable questions, and score the gap.
If you'd rather not do this yourself, Seer runs this as an engagement: we build the canon from your official sources, derive the tiered claim set, and score how every major AI platform describes you against it. Talk to us to learn more.
