The method, in four questions
How AI recommendations really work
“GEO,” “AI visibility,” “AI SEO” — under the jargon, one question: when a customer asks AI who to hire, does it say your name? Here is what actually decides that, in plain terms.
01What is this, really?
When someone asks ChatGPT or Perplexity for a plumber, the AI composes an answer naming three or four businesses — and stops. There is no page two, no position eleven. You are in the answer or you are not. “AI visibility” is nothing more mysterious than how often that answer includes you.
02How does the platform decide?
Two machines produce answers. For current questions the AI searches the live web and composes from what comes back — and what comes back is mostly not your website. In our measurement of 100 local businesses, Yelp appeared in more answers than any other source; Perplexity leans on Reddit, Gemini reads state license registries, and the winning businesses were readable in both places — their own site and the directories. The second machine is the model's trained memory, which updates on the platform's schedule, not yours.
One more thing most owners don't know: about 8% of small-business sites block AI crawlers outright — which can restrict the named crawler’s access. This does not establish that every AI search system is unable to find the business; other crawlers, indexes and sources can still supply information.
03What can you actually affect?
The inputs to the live-search machine: a site the crawlers can read, an identity that agrees with itself everywhere (name, phone, hours, services), pages that actually answer the questions customers ask, and presence on the directories your trade's platforms read. None of it is instant — we measured the lag ourselves, and even a directly-submitted page takes days to enter the indexes the AI platforms search.
04What can't you affect?
There is no submission form and no way to buy placement — and no technique has a demonstrated, guaranteed effect on what an AI says. Anyone promising a specific AI answer is guessing. The answers are also stickier than the daily-chart industry implies: 77.5% of business/question/platform combinations were strictly always- or never-mentioned in the cited study’s repeat runs. That observation is specific to its sample and settings; it does not establish how every business changes over time or the effect of an intervention.
∴A clear plan, with the evidence behind it
The paid assessment combines what AI says about your business with an inspection of your public website. We explain what we found, recommend the changes to work on first, and prepare replacement copy or business information where the evidence supports it. Each proposed change links back to the page or finding behind it. You review the work before putting it live.
- Your website. We inspect up to 30 linked public pages, prioritising services, locations, contact information and other useful content. Checks cover titles, headings, search descriptions, instructions about indexing, links and structured business information in the initial HTML. We also read your homepage after JavaScript loads, recording that browser evidence separately so existing business markup is not mistaken for missing code.
- The businesses AI named. We inspect up to five linked public pages on each of three confirmed competitor websites. We also save available directory observations, including business details and review data, with their sources and dates. Differences suggest work to consider; they do not prove why AI recommended a business.
- What changed. We keep the website findings separate from changes in AI answers, so an improvement on your site is not presented as proof that it caused a recommendation.
For the AI measurement, we ask each question repeatedly across the configured provider APIs — asked once, a miss could mean almost anything; asked fifteen times, it means something — publish the rate with its 95% range, and freeze your questions on the first run so a re-check measures a change in the answers, not a change in the questions. These API results do not represent every answer a customer may see.
Every report lists the pages checked and any limits. We retry recoverable retrieval failures; anything still unreadable stays unverified, not missing. The inspection covers initial HTML on linked public pages and a separate browser check of your homepage. Interior pages and competitor pages are checked through initial HTML; the report labels the method and date. It does not audit private plugin settings, analytics accounts, form delivery or actual search-index inclusion. A browser observation does not establish what every search or AI crawler receives. No assessment can guarantee placement, customers or a particular result.
The instrument itself, measured: sensitivity 85% (95% 79–90%) against an independent adjudicator across 60 businesses; platform failures excluded from rates rather than counted as absences; every published correction stays published. The full protocol ships inside every report, and the deep studies live on the research page.