Signals
We asked AI to audit our schema. It failed us on the two things we do best.
By Arnav Mukherjee, founder of TofuBofu · August 1, 2026
I asked Gemini a simple question about my own company: does tofubofu.com have its schema set up correctly for AI Overviews, AEO, and GEO? It came back with a confident, well-formatted, four-part verdict, complete with citations to our own pages. Two of the four verdicts were fair. The other two were flatly, checkably wrong, and they happened to be the two things we do most heavily.
I build an AI visibility tool, so I found this less annoying than perfect. What just happened to my own website is the exact problem the whole product exists to catch: an AI stating something confident, cited, and false about a company. So here is the case study, with the real answer, because it is the clearest argument I have ever seen for why you cannot trust an AI's description of a business, including your own.
What the AI said
Here is Gemini's scorecard, verbatim in substance:
Google AI Overviews readiness
Answer-first structure, punchy summaries at the top of each section. Fair.
GEO setup
Comparison pages, author bylines, publication dates, clear attribution. Fair.
AEO / FAQPage markup
Claimed we 'lean away from hardcoding rigid JSON-LD FAQ blocks' and rely on typography instead.
Organization schema + sameAs
Claimed we are 'lacking heavy traditional organization schema linking to verified social profiles via sameAs'.
It then offered, helpfully, to generate the exact Organization schema code we need to fix the entity confusion. Which would be a generous gesture if we did not already have it.
What is actually true
I checked the live site, view-source, the boring way. The two failing grades are the two things we use most:
What the AI claimed vs what the page actually serves
Every marketing and blog page ships FAQPage JSON-LD with six to ten structured question-and-answer nodes. Organization and WebSite schema are site-wide. The homepage and pricing carry SoftwareApplication and Offer schema. Content pages carry Article, a Person byline, and BreadcrumbList. And the sameAs entity graph links eight verified listings. The AI did not just miss these. It asserted the opposite, and offered to sell us the fix.
How it got there, and why it matters
The tell is in its own reasoning. It cited our "How it works" page, which describes what TofuBofu scans for on a customer's site, Organization, FAQ, Service, LocalBusiness, and then flipped that into a claim that we approach our own site differently. It took accurate text about one thing and confidently misattributed it to another. The result was fluent, formatted, cited, and false.
This is the entire reason AI visibility is a category. If an engine can get the basic, checkable facts about my own website wrong, with citations, then it can get the facts about your company wrong too, and it will say them to your buyers with the same confidence. You cannot manage that by asking an AI to summarize how you look. You manage it by reading the real answers the engines give, finding where they are wrong or where they name someone else, and fixing the signals behind them. Guessing is how you end up invisible, or misdescribed, without knowing it.
So we did two things with this. We made our own entity signals even harder to misread, an About page, Person schema for the founder, SoftwareApplication on the homepage, a richer sameAs. And we wrote this, because a confident, wrong AI audit of our own site is the most honest demonstration of the problem we could ask for.
Stop guessing what AI says about you
Run a free scan and read the actual answers ChatGPT, Gemini, Perplexity and the rest give when a buyer asks for a company like yours.
Get your free auditFrequently asked questions
Can AI accurately describe what schema a website uses?
Not reliably. In our own case, Gemini audited tofubofu.com and confidently reported that we lack FAQPage schema and Organization schema, when in fact both are present on every page. AI engines synthesize a plausible-sounding answer from what they read, and they can misread context and state the opposite of the truth with full confidence. If you want to know what schema or what AI answers describe your company, you have to check the real output, not trust an AI's summary of it.
Why did Gemini say TofuBofu lacks schema when it does not?
It appears to have read our product description, which explains that TofuBofu scans a customer's site for Organization, FAQ, Service and LocalBusiness schema, and flipped that into a claim about our own site, inventing the idea that we approach our own visibility differently. This is a classic AI failure mode: it took accurate source text about one thing (what the product does for clients) and confidently misattributed it to another (whether we use schema ourselves). The output reads authoritative and is simply wrong.
Does schema markup actually matter for AI visibility?
Yes, though less as a magic switch than as one signal among several. Structured data like Organization, FAQPage and SoftwareApplication helps engines understand and cite a business, and a rich sameAs entity graph helps Google's Knowledge Graph recognize you as a real brand. It is not the only thing that matters, answer-first content and third-party corroboration matter at least as much, but it is a cheap, foundational signal, which is why it is worth getting right and worth checking that it is actually present.
How do you check what AI actually says about your company?
You ask the engines directly, the way a buyer would, and read the real answers, rather than trusting any single AI's confident summary. That is exactly what TofuBofu does: it runs real buying-intent queries across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and Bing Copilot, and shows you the actual text of each answer, where you are named and where a competitor is named instead. The whole point is that an AI's confident description, including of your own site, is not evidence. The real answers are.
What schema does tofubofu.com actually use?
Organization and WebSite schema site-wide, SoftwareApplication and Offer schema on pricing and the homepage, FAQPage JSON-LD on every marketing and blog page, Article, Person byline and BreadcrumbList schema on content, and a sameAs entity graph linking our verified listings including Capterra, the MCP Registry and mcp.so. In other words, the two things the AI audit said we were missing are the two things we use most heavily.
Is this a reason to distrust AI search entirely?
No, it is a reason to instrument it. AI search is where a growing share of buyers now start, so being named there matters more each month. The lesson is not to dismiss AI answers but to stop guessing what they say. An AI got the basic, checkable facts about our own website wrong with total confidence, which is the clearest possible argument for measuring the real answers about your company rather than assuming, and for building the signals that make those answers accurate.