How the score has evolved
We version our scoring engine in the open and stamp every result with the engine that produced it. No black box, no quiet rewrites. This is the honest record of what changed and why. Light on detail by design, you can read the full methodology any time.
-
v2.9.3currentA comparison check that counted the word inside other words
Two corrections to the sub-question coverage added in the previous release, both found by measuring it against real sites rather than by reasoning about it. First, one of the comparison terms was short enough to match inside ordinary words, so a page reading "our TVs are excellent quality" was counted as answering how the business compares to its alternatives. A page with no comparison content at all could be marked as covered. That term is removed rather than patched, because the four remaining ones carry the meaning and a three-letter match is not worth what it invites. Second, several sub-questions were phrased so literally that pages answering them in ordinary English went uncounted: a section headed for who this is built for was invisible unless it used one specific wording. The vocabulary now covers how a buyer would actually write it. Scores may rise slightly where a site already answered a question in its own words, and may fall where a comparison was counted that was never there. 41 scored signals.
-
v2.9.2Names the off-site sources it checked, and therefore the ones it did not
The off-site section reported a result without saying which sources produced it, so a business reading that no independent record was found had no way to know that review sites and business directories were never checked at all. Silence about what was skipped reads as coverage. The report now lists the sources it queried directly and states plainly that anything blocking automated checks was not among them, so nothing here should be read as a verdict on those. Two sources were also added where they mean something and gated to the business type given: the US healthcare provider registry, which is close to a census for a healthcare organisation rather than a directory anyone opts into, and the open book catalogue for authors and publishers. Both are reported and neither is scored, on the same basis as the rest of this section. 41 scored signals.
-
v2.9.1Adds the one off-site gap you can actually fix yourself
The off-site check now also looks at OpenStreetMap, the open map data redistributed into a great many apps and assistants. It is reported separately from the encyclopedic sources and worded differently on purpose. A missing encyclopedia entry is context: most businesses do not meet the notability bar and writing your own is a conflict of interest that usually gets it removed, so we say so and score nothing. OpenStreetMap is the opposite. Anyone may legitimately add a real business, it is free, it takes about ten minutes, and it is handed over as an instruction rather than an observation. The check only runs when a local market was given, because a business that did not name one is not one we can assume has premises, and telling an online-only brand to add itself to a public map would be asking it to put something there that does not physically exist. Matching is exact on the map name, and similarly named places are counted as rejected rather than counted as you: the geocoder is fuzzy enough that a loose match would report almost every business as already listed. The wording is conditional on purpose: a business that names a local market has not necessarily got premises the public can visit, so the finding opens with that condition and says plainly that an online-only business is not marked down either way. Still not scored. 41 scored signals.
-
v2.9.0Scores whether you answer the questions AI actually asks
An assistant does not run the question it was asked. It rewrites one buying question into roughly a dozen related sub-questions, runs each, and assembles its answer from whichever pages answer each one, retrieving passages rather than whole pages. That is the mechanic behind the thing businesses find hardest to explain: a site can rank first and still be absent from the answer, because ranking is decided page by page and citation is decided sub-question by sub-question. The report now measures how many of those sub-questions your site actually answers, and names every one it does not, in the words a buyer would use. Credit is proportional, and a page with no headed sections is excluded rather than marked zero. This pairs with the passage check added earlier and measures the opposite failure: that one asks whether your sections can be quoted at all, this asks whether they answer what was asked, and a site can pass the first while failing the second. The sub-question set is fixed rather than generated, so the same site always produces the same reading and a score cannot be moved by re-running it. Scores will move, and will fall for sites that cover few of these questions. That is the finding, not a side effect. 41 scored signals.
-
v2.8.0Looks at whether anything off your site actually carries you
Roughly 84% of AI citations point at third-party media rather than at pages a business owns, and until now this engine measured that entire half with a single on-page check: does your site link out to an authoritative source. That check reads your page. It cannot tell you whether anything is at the end of the trail, and its wording implied otherwise, so it now says plainly that it checks your page and not the world. Alongside it the report now goes and looks, querying the public entity databases for a record under your business name. That result is REPORTED AND DELIBERATELY NOT SCORED. Most legitimate businesses do not meet the notability bar for an encyclopedic entry and never will, and writing your own is a conflict of interest that usually gets it removed, so scoring its absence would mark an honest business down for failing to do something it cannot honestly do. That is the same fault corrected on the Trust pillar in the previous release. When the databases do not respond the result is recorded as unmeasured rather than as an absence, and it affects nothing. No signal was reweighted and no score moves as a result of this release. 40 scored signals.
-
v2.7.0Stops asking businesses to republish ratings that are not theirs
Two corrections to the Trust pillar, both found on a real author’s report. First, a missing AggregateRating was always scored as a gap, and the recommended fix was to publish one. For a business whose ratings were earned on someone else’s platform, doing that means republishing another company’s data as your own structured data, which breaches that platform’s terms. The engine was recommending a violation and scoring the honest choice as a failure, on the pillar named Trust. When reviews are visible on the page and the platform they came from is linked, those two signals are now marked not applicable and excluded from the score, exactly as Local Presence already is for a business with no premises, and the report says why instead of claiming the signal could not be observed. Second, the review-platform check knew Google, Yelp, Trustpilot, BBB and Facebook and nothing else, so every author, publisher and product brand was told no review platforms were linked while linking the one their reputation actually lives on. Amazon, Goodreads, Barnes & Noble and Apple Books now count. Scores may rise on re-evaluation for businesses whose reviews live off-site. No threshold was relaxed and no signal was reweighted. 40 scored signals.
-
v2.6.1Corrects two passage checks that under-scored real writing
Two detection faults in the passage check, both of which were penalising sections that were already correct. First, a subject named by an acronym shorter than four letters was invisible to the check. A word-length filter meant to discard filler words was also discarding AEO, GEO, SEO, RAG, LLM and API, so a page whose whole subject was one of those had no detected subject at all, and its sections were failed for not naming a subject they in fact named in their first line. Second, a section under a yes or no question that opened with the single word "No." was marked as a fragment. That is the strongest possible opening for a passage an answer engine will quote whole, and it was scored as though it said nothing; a bare answer with no substance behind it is still marked down. Both faults moved scores in the same direction, downward, on exactly the kind of direct, plainly written content this check exists to reward. Scores may rise on re-evaluation. No signal was added, removed or reweighted, and no threshold was relaxed. 40 scored signals.
-
v2.6.0Scores the passage, not just the page
AI systems do not retrieve your page, they retrieve a slice of it. Your content is split into self-contained passages that are scored and cited independently of the rest of the document, which is the real reason a strong page gets passed over while a weaker one gets quoted: the weaker page had one section that made sense on its own. The check now reads every section as its own passage and asks whether it survives being pulled out and read cold. A section that opens by pointing at the paragraph above it, that never names its own subject, that poses a question and then winds up before answering it, or that runs past 600 words unbroken, is a section no answer engine can quote. Credit is proportional: this scores the share of your sections that stand alone, not a pass or fail on the whole page, and a page with no sectioned writing is excluded rather than marked zero. Card grids and navigation labels are excluded for the same reason. This also brings the check in line with Google’s own guidance, published in May 2026, that its systems extract the relevant passage from a page themselves and that authors should not pre-fragment their writing to help: the work is making each section coherent on its own, not chopping the page up. 40 scored signals.
-
v2.5.0Scores only what can be corroborated
Two corrections, both of which move scores. First, llms.txt is no longer scored. Server-log studies published in June 2026 found that 97% of published llms.txt files received no AI-crawler requests at all, and Google confirmed the file is not required. We still check it and report it, but crediting a file the engines demonstrably do not read would inflate a number we promise is real. Second, sameAs entity links are now graded by who controls the target rather than by whether the field exists: links to an identifier maintained by someone else (a Wikidata entry, a company registry, a review platform) count fully, social profiles you publish yourself count partially, and links pointing only back at your own domains no longer count at all. The weight released by llms.txt moved to signals that can be verified: structured data and facts readable as plain text. 39 scored signals.
-
v2.4.0Reads what AI can actually read
Added two signals reflecting how AI answer engines fetch the web: whether your key facts (prices, numbers) are readable as plain text rather than locked behind JavaScript, and whether your site exposes a path to a third party that can corroborate you. 40 signals total.
-
v2.3.1Broader signal coverage
Expanded the check to 40 signals for a more complete read of how AI systems see a site.
-
v2.3.0Verified badge and public reports
Every check now has a shareable report page, and sites that score well can display a verified AI Visibility badge.
-
v2.2.2Compare locations
Added a multi-location comparison, powered by AIOStory, so you can see how a brand's story holds across its locations.
-
v2.2.1Clearer pillars and navigation
Standardized the six pillars and simplified the site so the results read clearly.
-
v2.2.0Detects when AI cannot read a site
Added detection for sites blocked from AI crawlers, the highest impact issue to catch.
-
v2.1.1Honest usage signals
Added a real, unpadded activity counter and clearer notes on how scoring works.
-
v2.1.0Trust and Local added
Added the Trust and Local Presence pillars for a fuller picture of a business.
-
v2.0.0Real deterministic engine
Rebuilt scoring on a real, reproducible engine, so the same site always gets the same score.
-
v1.0.0Launch
The first public AIOInsights Trust Visibility Check.
Want to see it in action? Run a free AI visibility check.