The 62% Number Is Not the Finding

A study comparing Google rankings against ChatGPT visibility found that brands ranking on Google’s first page were named in ChatGPT answers 62% of the time. The number traveled fast. It has been quoted in vendor decks, conference slides, and roughly every pitch for a generative engine optimization retainer written since.

The 62% is real. It is also the least useful number in the study.

What the study actually measured

The headline comparison: page-one Google presence produced a ChatGPT mention 62% of the time. Underneath that, four findings that got far less attention.

The correlation between Google rank position and ChatGPT position was 0.034. Not weak. Effectively zero. Even when a brand appeared in both places, where it ranked on Google told you almost nothing about where it landed in the AI answer.

Turning on browsing in ChatGPT moved alignment with Google by one percentage point. The model reaching out to live search results barely changed which brands it named.

Query intent did not matter. Exploratory queries, feature-comparison queries, and brand-seeking queries all landed between 61% and 63% overlap. Rewriting content to sound more like the questions people ask an assistant did not appear to be the variable.

Brand-level variance was wide. Coursera and GoDaddy cleared 83% overlap. Hostinger sat between 32% and 34%. Same categories, same search engine, radically different results.

The part that changes what you do on Monday

If rank position carried real predictive weight, the correct response would be to keep doing SEO and wait for the benefit to arrive. If intent type mattered, the correct response would be to rewrite your content in question-and-answer format, which is what most of the tooling market is selling.

Neither holds. A 0.034 correlation and a flat result across intent types point somewhere else: the variable that decides whether a model names you is not on the page you are optimizing. It is in whether the model has a clear, corroborated understanding of what you are and who you serve, assembled from sources you mostly do not own.

That is a different problem with a different sequence of fixes. It is also, for most firms, a measurement problem before it is a content problem.

The failure mode underneath the study

The study compared brands that models already understood well enough to place in a category. That is a generous baseline. In audit work, the more common failure sits a level below it.

Consider a firm whose name contains a term that means something specific in an unrelated industry. Every retrieval system in the chain reads it the same wrong way. Google ranks it for terms in the wrong category. AI assistants answer questions about the wrong category. When someone asks for the best firm of the type it actually is, it does not appear at all, while fifteen competitors do.

Nothing on that site is broken. Load time is good. Rankings exist. Pages are indexed. The firm is invisible for the only questions worth being visible for, because the machines never established what it was.

For a company in that position, 62% is not the risk. Zero is the risk. And the gap does not show up in any dashboard the marketing team currently reads, because rank tracking reports the terms you rank for, not the category you failed to claim.

Three decisions, not a new playbook

The vendor framing says AI search requires a separate discipline with its own budget line. That framing sells software. It also skips the part where most organizations cannot answer three questions about their own site.

What do you want to be retrieved for? Not keywords. Category. If an assistant had to describe your firm in one sentence to someone who had never heard of you, what is the sentence, and does anything on the public internet support it? Ranking is a positional question. Retrieval is a categorical one, and the categorical question gets answered first.

What on your site is actually yours? Many firms on marketing platforms publish a syndicated content library shipped to every customer of that platform. The same articles sit at the same URL paths on dozens of competitor sites. A retrieval system deduplicating near-identical documents has no reason to attribute any of it to you. Volume is not evidence. Original publishing is.

What counts as a lead? A large share of the sites we audit have analytics installed and zero conversion events configured. Those organizations cannot evaluate AI visibility because they cannot evaluate any visibility. Adding a GEO tracking tool to that stack produces one more number nobody can act on.

What to do this quarter

Fix measurement first. Configure conversion events that reflect real intent. Instrument the assisted paths, including direct and branded traffic, because AI-referred visitors frequently arrive with no referrer at all.

Then run the prompts. Take the twenty questions a qualified buyer would actually ask an assistant before they call anyone, run them across ChatGPT, Claude, and Gemini, and record who gets named. It takes an afternoon and it is a better diagnostic than most reporting suites, because it tells you the category the models put you in.

Then fix the machine-readable layer: entity markup that says what you are, schema on the pages that carry your expertise, an internal structure that connects your original work to the terms you want to own.

Then, and only then, decide what to publish.

That order matters more than any single tactic in it. Ranking well and being retrieved are not the same outcome, and the study is a useful reminder of that. What it does not say is that you need a second team, a second budget, and a second set of tools. It says the thing you have been measuring was never the thing that decided whether you get named.

If your search reporting looks healthy and your pipeline does not, that gap is worth an afternoon of diagnosis before it is worth a retainer.

Book the Assessment

Brett Berchtold is Principal of Berchtold and a two-time Sitecore MVP in Digital Strategy. Source: Search Engine Land, September 2025.