A client forwards a screenshot. They asked ChatGPT for a good accountant in their town, three firms came back, and none of them was theirs. Then comes the obvious question: how do we get in there. The honest first answer is that there is no ranking to buy, no console to log into, and no reliable way to check your position. There are still several things that demonstrably change the odds.
How AI assistants pick which businesses to name
Two different mechanisms are at work, and confusing them produces most of the bad advice on this topic.
The first is retrieval. When an assistant has search enabled it runs one or more queries, retrieves a handful of pages, and writes an answer grounded in what it read. This is how nearly every local question gets handled, because the model knows perfectly well it has no current information about which accountancy firm in Shrewsbury is any good. The output depends almost entirely on which pages came back.
The second is parametric memory: what the model absorbed during training. That is where enormous, long-established brands get named without any retrieval happening. For a local business it is effectively unavailable. You cannot get into training data on any timescale that matters, and anyone selling you that is selling you nothing.
So the practical target is retrieval. Be present, clearly, in the sources that get pulled and summarised when somebody asks the question in your category.
One consequence is worth internalising early. Assistants disproportionately name whatever appears on a page that already lists several businesses. A model asked for three good options finds a page containing three good options and paraphrases it. That page is very often not any of those businesses' own websites.
The sources these models lean on
| Assistant | Retrieval layer | What that means locally |
|---|---|---|
| ChatGPT with search | Its own index plus Bing-derived results | Bing visibility and directory presence matter more than most SEOs assume |
| Gemini and Google AI Mode | Google Search and the Maps corpus | Your GBP feeds it directly |
| Perplexity | Its own crawl plus live search | Heavily citation-led; roundups and directories dominate |
| Copilot | Bing | A Bing Places listing is worth the twenty minutes |
Across all of them the same page types keep appearing: "best X in Y" roundups, established directories, professional body member lists, local press, and community threads. Reddit gets cited far more often than its traffic share would suggest, because it is where genuine unfiltered opinion about local services lives.
In practice this makes the highest-value work unglamorous. Being listed accurately on a trade association member directory is worth more here than another blog post. Being named in the local paper's roundup of the town's best cafés is worth more than both.
Why your GBP still does the heavy lifting
It is tempting to treat AI search as a separate discipline. Mostly it is not. The profile does three things that feed directly into how a model describes a business.
- It is the canonical fact source. Name, address, phone, hours, categories, service area. When an assistant states your opening hours, that number came from a listing, and Google's is the most widely mirrored listing in existence.
- It resolves the entity. A model needs to know that "Baxter Plumbing", "Baxter Plumbing Ltd" and "Baxter's" are one business. Consistent naming across the profile, the site and every directory is what makes that resolution reliable. Inconsistency does not get you penalised so much as left out, because the model cannot confidently attach the good things it read to the business being asked about.
- It supplies the sentiment. Review volume and review text are the closest thing to a quality signal a model has for a local business. When an answer says a firm is "known for being responsive with smaller commercial clients", that came from reviews saying so.
None of this is new work. It is the same profile hygiene that mattered before, now with a second consumer reading it.
Structured information that machines can read
Retrieval-grounded answers are built from pages. The easier a page is to parse, the more likely a specific claim from it survives into the summary.
- LocalBusiness schema, or the more specific subtype where one exists, with address, geo, opening hours, telephone and areaServed. Add sameAs pointing at your GBP, your main directory listings and your social profiles. That is the machine-readable version of "these are all the same business".
- One page per service per location, with the service and the place named in the H1 the way a person would say it out loud. Not "Our Solutions".
- Answers before preamble. Put the direct answer in the first two sentences under the heading. Retrieval works on passages, and a passage that answers the question outright is more useful than one that builds towards it.
- Real FAQ content, phrased as the question a customer actually asks, answered in fifty to eighty words. FAQPage markup no longer earns rich results for most sites, but the structure still helps a passage stand on its own.
- Facts in text, not only in images. Hours in a graphic, price lists inside a PDF scan, service areas shown on a map with no written list - all invisible.
Measuring something you cannot rank-track
There is no position to track. Answers vary by phrasing, by user, by session and by whatever the model happened to retrieve in that moment. Any tool claiming a stable AI ranking is reporting one sample of a noisy process.
What does work is a prompt panel. Write fifteen to twenty prompts a real customer might type - "best [service] in [town]", "who should I call for [problem] near [area]", "is [client name] any good" - and run them once a month across the two or three assistants that matter for that client. Log three things: whether the business was named, roughly where in the list it appeared, and what the assistant said about it. That third column is the valuable one, because it tells you which sources the model is reading, and it is frequently a page the client had forgotten existed.
Alongside that, check referral traffic. Analytics will show sessions referred from chatgpt.com, perplexity.ai and similar. Volumes are usually small and intent is usually high. Search Console will not separate AI-surfaced impressions in any usable way, so do not build a report that depends on it.
Ampli5 Pulse is not a rank tracker and does not monitor what assistants say about anyone. What it covers is the downstream half: calls, direction requests, website clicks and views per location, alongside Search Console data, with white-label reports if you are presenting this to clients. Pair that with a manual prompt panel and you have something defensible for a monthly review, which is more than most agencies currently bring to this conversation.
What is not worth your time
- Paying anyone for "ChatGPT optimisation" who promises placement. There is no placement to buy and no mechanism through which to sell it.
- Keyword stuffing for AI. Retrieval works on passages and meaning. Density does nothing here that it does not already fail to do in Search.
- Mass-producing thin content in the hope of being cited. The pages that get cited are the ones that already rank or that aggregate real opinion. Volume without either is invisible.
- Chasing every new assistant. Pick the two the client's customers actually use, and check the rest twice a year.