What GeoScan measures, and what it admits it cannot

Updated: 2026-08-29

Quick answer: the engine runs 172 checks in 8 groups on every page. Ten of its thresholds come from published measurements. The other sixty-five variables are judgements informed by that research, and this page says which is which, because a tool that hides the difference is asking you to trust it rather than check it.

Which measurements does the engine rest on?

Every figure below comes from a study or from our own scans. Each row links to the page that explains what it means and what to do about it.

The first four come from GEO-bench, the Princeton benchmark that tested single content modifications against AI visibility. The overload figure comes from research on brand bias in language model recommendations.

Citability

Whether a model can find, read and use the page as a source. Content multiplied by access, not added to it.

Memorability

Whether a model knows the brand before any page is fetched. Returned as empty below three independent external signals.

Check

One separate verdict inside the report. Not a group and not a signal source: a single yes or no.

What did the engine find on our own site?

We publish our own numbers because a vendor that does not is asking you to go first.

Four models were asked six customer-style questions each. None named us. The full account is in Our site scored 96 and got zero citations, and the reason we report two numbers instead of one is in One score hides the problem.

What are the 8 groups?

Seven groups describe the page. The eighth describes the brand, and it is the one you cannot fix by editing HTML.

Structured data is the only group whose count moves: the set of schema types expected of a law firm is larger than the set expected of a one-page site. That is why the total is stated as 172 rather than as a single fixed number.

How is the score built?

Two numbers, not one.

Citability answers whether a model can use the page. It is calculated as content × (0.6 + 0.4 × access). Multiplication, not addition: a site cannot collect points from clean infrastructure while having nothing worth quoting. Nine conditions zero it outright, because a page that cannot be fetched has no partial credit to award.

Memorability answers whether a model has any prior about the brand. It is computed only when at least three independent external signals are available, and returns nothing at all below that. A low number would mean "we looked and found little"; the absence of a number means "we did not gather enough to have an opinion". Collapsing those two is how invented measurements get into reports.

What do we not measure, and will not pretend to?

Your share of AI answers. Nobody can measure it. Conversations with assistants are private, query volumes are unpublished, and every percentage of "AI share of voice" on the market rests on a denominator that does not exist. The argument is in full in The denominator problem.

Whether a specific change caused a specific citation. We cannot isolate a variable in a system we cannot observe. We can measure whether four models name you before and after, which is a smaller claim and a true one.

Four things we watch but score at zero. An llms.txt file, a markdown twin of each page, content negotiation for markdown, and an MCP endpoint. All plausible, none adopted by any major provider, so all weighted zero and reported separately. The reasoning, and what would change it, is in We gave llms.txt a weight of zero.

What are the honest limits of this model?

Ten thresholds of seventy-five come from published measurements. The rest are judgements: informed by the research, argued in the pages linked here, and not validated against outcomes.

That makes the current engine a better-grounded hypothesis than its predecessor, not a validated instrument. The previous version scored our own site 96 out of 100 while four models returned zero mentions, which is what a confident number looks like when it measures the wrong thing.

Validation is the work we have not finished: accumulating scans and correlating citability with what the jury of models actually says. Until that exists, this page is the closest thing to a warranty we can honestly offer.

Frequently asked questions

Why 172 and not a round number? Because it is counted from the model rather than chosen for the landing page. Every check that produces its own verdict is counted once; thresholds of a single measured value are not counted as several. The previous engine reported 305 and that was true for the previous engine.

Does a higher score mean more citations? Not on its own. It means the page can be used if a model reaches for it. Whether it reaches depends on memorability, which moves through activity outside your site.

What is a blocker? A condition that makes citation impossible rather than unlikely: noindex, robots.txt closing the site, a captcha served to bots. Nine of them, and each drops citability to zero instead of shaving off points.

Why does memorability sometimes come back empty? Because fewer than three independent external signals were found. Reporting a low number there would be inventing a measurement.

Can I see every check? Yes. The report lists all 172 with what was expected and what was found, grouped checks naming each blocked bot and each missing tag.

How often does the model change? Rarely, and every change is recorded in the version history with the reason.