Schema won't save you: what 10,000 local businesses taught us about AI visibility

If you've read anything about "AI SEO" in the last few months, you've probably met the same checklist five times over. Add JSON-LD for Article, FAQ, HowTo and Speakable. Publish an llms.txt file. Get your @id and @graph knowledge-graph markup in order. Do this, and ChatGPT will finally notice you.

It's delivered with total confidence, usually as a numbered list, and almost never with any data behind it. Nobody seems to have actually checked whether any of it works.

So we did. We audited 10,000 real US local businesses, re-ran them through a clean environment, and measured what actually correlated with being found, liked and recommended by ChatGPT and Perplexity. Some of the current GEO playbook holds up. Most of it, on our numbers, barely moves the needle.

The playbook, summarised

To be fair to the genre, the standard advice isn't unreasonable on its face. If AI tools are reading the web to answer questions, it seems sensible that machine-readable markup would help them read it better. The playbook that's circulating right now generally recommends:

Crawlability basics, so bots can actually reach your content. JSON-LD structured data across Article, FAQ, HowTo and Speakable schema types. Knowledge-graph markup (@id, @graph) to help AI "understand" who you are as an entity. An llms.txt file, the new-ish idea of writing a set of instructions specifically for AI crawlers.

It's a coherent theory. It's just not what we found when we checked it against 10,000 real businesses and real ChatGPT and Perplexity outputs.

What we did instead of guessing

For our AI Visibility Report, we took a sample of 10,000 US local businesses and re-ran all of them through the Insites platform in a single, standardised audit environment, so every business was measured against identical criteria. We then queried ChatGPT and Perplexity about each one across three layers: whether AI knows the business exists and gets its details right (awareness), whether AI speaks positively about it (reputation), and whether AI actually recommends it for relevant local searches (recommendation). Every AI-readiness and technical SEO factor was then correlated against those outcomes. Full methodology and the complete data set are in the AI Visibility Report, and we've broken the three layers down in full in the AI visibility benchmark.

The needle that didn't move

Here's where it gets uncomfortable for anyone who's built a service line around AI-readiness markup.

Take FAQ schema. Businesses ChatGPT was aware of were slightly more likely to have it (9.4% vs 5.6%) but when we looked specifically at whether it influenced recommendation, the relationship disappeared entirely.

llms.txt is the most quotable one, and it's worth its own look: we go deeper on it specifically in why AI-optimisation tricks like llms.txt barely move the needle. The short version: adoption was a bit higher among businesses AI surfaced (29.2% vs 24.7%), but that gap is weak, and plenty of businesses that dutifully added an llms.txt file are still nowhere to be found in ChatGPT or Perplexity's answers. Writing a file that only crawlers read turns out not to be a substitute for actually being a well-established business.

And the numbers on missing structured data are the real tell: even businesses AI is happily recommending are, on average, still missing over half of the schema types considered "recommended." If schema were the gatekeeper the playbook implies, that shouldn't be possible.

What actually separates winners from losers

While AI-readiness markup nudges things by a few points, these signals move the needle by tens of points.

The single strongest signal in our entire data set: businesses found by both ChatGPT and Perplexity had an average of 133.4 Google reviews. Businesses found by neither had 10.7.

That's not a rounding error, it's more than twelve times the review volume, and we've pulled that thread on its own here. And it wasn't a one-off. A cluster of local and reputation signals told the same story throughout the report:

Appearance in the Google Local Pack ran at 43.4% for businesses found by both AI platforms versus 10.1% for those found by neither. Monthly organic traffic averaged 29,044 for found businesses against 661.8 for the rest. Businesses ranking for their own target keywords hit 71.1%, compared with 13.5% for those AI ignored. Content depth mattered too: businesses AI surfaced had an average of 34.8 pages on their site, more than double the 15.2 pages on sites AI didn't find, and their content was noticeably fresher.

Recommendation (not just awareness) told a similar story. Review volume nearly doubled between businesses that were recommended and those that were never mentioned, and Local Pack presence was the single strongest predictor: 74.6% for highly recommended businesses versus 21.1% for those never recommended.

None of this is exotic. It's the same fundamentals that mattered for local SEO before anyone had heard of ChatGPT.

Why this makes sense

The framing that clicked for us while going through the data: AI systems aren't scoring your markup. They're asking whether there's enough corroborating evidence across the web that a business is real, active and trustworthy.

Schema is one business telling an AI model what it is. A hundred reviews, a claimed and photo-filled Google Business Profile, consistent listings across a dozen directories and a site that ranks for its own name are a hundred other sources telling the AI model the same thing independently. One of those is much easier to fake, or simply to get wrong, than the other. AI systems seem to have worked that out, whether by design or by training data, and they weight accordingly.

The honest caveat

Technical SEO and AI-readiness markup aren't useless, they're a baseline. Missing page titles showed one of the strongest correlations we measured with AI awareness (0.9% missing titles among AI-found businesses versus 3.5% among those AI missed), so basic hygiene clearly matters. But the pattern stops there: going from no schema to imperfect-but-present schema seems to help a little, while going from imperfect schema to textbook-perfect schema barely registers. Going from 10 reviews to 130, by contrast, is one of the biggest levers in the entire study.

What to actually prioritise this quarter

If you're advising local businesses on AI visibility, or you're an SMB owner who's just been sold a schema package, the higher-leverage list looks like this: claim and fully complete the Google Business Profile, including photos, categories and verification. Build a genuine, steady flow of reviews rather than a one-off push. Fix inconsistent NAP data across directories and listings. Keep the website updated with real, current content rather than a static "set and forget" build. Build out actual content depth: more pages, more services covered, more reasons for AI to understand what the business does.

None of that is glamorous. It's also what was working before AI search existed, which is probably the point.

Where to go from here

If you want to see how a specific business stacks up against these exact signals, Insites' free AI visibility audit runs the same checks used in this research. For the full data set, methodology and every correlation we measured, the complete write-up is in the AI Visibility Report.