An AI model answers with what it can extract, attributes what it can identify, and skips what it can't parse. So showing up comes down to three jobs: Give it text worth quoting, an identity it can't confuse with anyone else's, and pages shaped like the questions people actually ask. None of that is exotic. Most of it is publishing discipline that was always good practice, now with a new reader.
The part that is genuinely new is that you can't watch it happen. More on that at the end, because it changes how you should spend.
What an AI model actually does with your website
When someone asks ChatGPT or Google's AI who in town does what you do, the model does something narrower than browsing. It pulls text it has read, decides what answers the question, and decides whose name to attach. Three failure points, three jobs.
Quotable text. AI models lift rendered words off pages, not hidden code. A question-shaped heading with a direct, self-contained answer under it is the exact shape they extract best, which is why this article opens by answering its own title. Pages that bury the answer under three paragraphs of wind-up give the model nothing to lift, and it quotes somebody who got to the point.
An unconfusable identity. The model has to be sure which business it's talking about before it will attach your name to anything. That certainty comes from corroboration, meaning the same name, address form, and web address appearing consistently across your site, your Google profile, your social accounts, and directories. Machines cross-check the way a skeptical lender would. Every inconsistency is a reason to hedge, and a hedging model just leaves you out.
Question-shaped pages. People ask AI models questions in plain words. Pages that answer named questions, in the phrasing a person would use, meet the query where it lives. That is a writing decision, made at the outline, not a tag you add later.
How badly we started, on our own site
We run this program on our own property, and the honest place to begin is what we found when we first measured, in mid-2026. Google's AI summary was confusing Pigasus Group with a planning consultancy in England called Pegasus. The site ranked around 22nd for its own name. And a set of national AI prompts about what we do mentioned us exactly zero times. A marketing consultancy, invisible to the surface it advises on, and being mistaken for a firm an ocean away. We publish that number because the fix is the demonstration.
What we changed, in order of how much it mattered. We rewrote the site's machine-readable files — including one called llms.txt, which is nothing more mysterious than a plain-text page of your key facts, written for AI models to read — because ours still described positioning we had abandoned and never mentioned the services we lead with. We said the Pegasus thing out loud where machines read: who we are, who we are not, spelled plainly. And we wired the identity corroboration properly, so the site, the social profiles, and the Chamber listing all vouch for one entity instead of four half-entities.
Did it move? The first re-check, three days after those fixes shipped, found the branded AI Overview describing the business accurately — right firm, right town, right services. The English planning firm was gone. Three days is faster than these systems usually re-learn, so we read it as directional rather than settled, and the fuller question set gets re-asked on a schedule. But the before and after are both on the record, dated, which is the only way this kind of claim deserves to be believed.
One thing we deliberately did not do is chase schema tricks. Structured markup helps machines confirm who you are, and it is worth doing right, but models quote the visible text. Markup describing content a reader can't see is testimony without a witness. If a choice ever comes down to better hidden code or a clearer visible answer, the answer wins.
The judgment call: Measuring a surface that hides from you
Here is the uncomfortable part. Search Console and analytics can show you Google queries and site visits. Neither can show you the conversation where someone asked an AI model for a recommendation and never touched a search results page. That conversation is invisible to every dashboard you own, and it is precisely the one this work targets.
The only instrument we have found honest is manual: a fixed set of questions, asked in the actual AI apps on a schedule, answers saved verbatim, compared over time. Slow, unglamorous, and real. The API versions of these models answer differently than the apps do, and the apps adjust for where you are asking from, which for a local business is the whole game — so a tool promising automated AI rank tracking is usually measuring a system adjacent to the one your customers use. Ask what exactly it queries before you pay for it.
Expect lag, too. Models retrain and refresh on their own calendars, weeks or months, so a change you ship today shows up in answers on their schedule, not yours. This is a quarterly read, not a daily scoreboard.
What this will not do
Being extractable and identifiable gets you accurately represented when the question is asked. It does not make the model like you. When someone asks who is good, models lean on the public evidence — reviews, ratings, how complete and current your profiles look — which means the rest of the baseline is what actually feeds this: the profile, the reviews, the consistent identity. AI visibility is what the other items compound into, read by a machine that quotes whatever the record supports, rather than a separate channel you bolt on at the end. The argument for taking the surface seriously hasn't changed since we made it; what we'd add now is that the work turns out to be ordinary, and the measurement turns out to be the hard part.
Start with one test that costs nothing: Open an AI app you don't use, ask it what your business does and who else in town does the same thing, and read what comes back as though a stranger wrote it. Whatever is wrong in that answer is your work list, in roughly priority order. Ours had a British planning firm in it. Yours will have something.