SEO or GEO: what should a business in Devon actually invest in?

What do SEO and GEO actually mean?

Search engine optimisation is the work of getting a page to rank in a list of links. Generative engine optimisation, sometimes called answer engine optimisation, is the work of getting a business retrieved, cited and named inside an answer a model writes. The first competes for a position. The second competes for a mention.

The mechanism matters more than the label. Google's own guidance names two things its generative features do: retrieval-augmented generation, also called grounding, and query fan-out (Google Search Central, 15 May 2026, updated 10 July 2026). Grounding means the model fetches live pages and writes its answer from what they say, rather than from memory. Query fan-out means one typed question becomes several separate searches behind the scenes.

Take a patient typing: “emergency dentist near Totnes who can see a child on Saturday and takes new NHS patients.” That is not one search. The assistant may issue four: same-day and emergency appointments in the area, Saturday opening hours, practices treating children, and which practices are accepting NHS registrations. It then assembles one answer naming two or three practices.

For a five-surgery practice, that changes what a website has to do. If the Saturday hours exist only in an image of the front door, and NHS status is implied by a friendly line about welcoming new faces, then two of the four sub-queries find nothing to retrieve. The practice may still rank perfectly well for “dentist Totnes”, because that is a different question being answered by a different system.

Scale is the reason this is not a niche concern. Google Search handles around 3 billion searches a month in the UK (Ofcom, Online Nation 2025), and a meaningful share of them now return a written answer rather than only a list.

Where they share foundations, and where they genuinely diverge

Google is explicit that AEO and GEO are still SEO, that generative features run on its existing Search index and ranking systems, and that a page must be indexed and eligible for a normal snippet before it can appear in an AI answer at all (Google Search Central). In other words, everything that makes a page findable in the conventional sense is a precondition, not an alternative.

That is the overlap, and it is real. It is also the limit of the overlap, for two reasons. First, Google's guidance describes Google Search. It says nothing about ChatGPT, Perplexity or Claude, which is where a large share of UK AI use now sits. Second, sharing an index does not mean sharing an outcome. Ranking is about a page. Being named in an answer is about a business.

Read the right-hand column and the problem becomes visible. None of it is measured by the left-hand column, and none of it fails loudly. A business can rank first and still never be named in an AI answer.

What does the AI-specific work actually consist of?

Four things, none of which appear in a ranking report.

Entity resolution. This means whether an assistant can reliably tell your business apart from similarly named ones, and then describe it accurately. Suppose a family law firm in Exeter shares most of its name with a larger firm in Manchester. An assistant asked for a solicitor in Devon may merge them, and answer with the Exeter firm's name attached to the Manchester firm's opening hours, partners and practice areas. The work here is consistent naming everywhere the business appears, organisation markup generated from the text visible on the page, verified external profiles, and unambiguous statements of town, county and country. A ranking report has no field for “described as a different company”.

Extractability. This means whether the facts a customer needs exist as findable, plainly worded text, rather than being implied, buried inside a general page, or held only in an image or a PDF. A builders' merchant near Newton Abbot may have a delivery radius, trade account terms and a cut-to-size service, all of them stated in a downloadable price list. Because of query fan-out, each of those facts may be retrieved by a different sub-query, and a PDF attachment is a poor place to be retrieved from. Brevity helps too: generated answers are short, and they draw on several sources at once. A fact stated in one plain sentence has a better chance of surviving into a short answer than the same fact spread over a paragraph.

Crawler accessibility. This means whether AI crawlers are permitted in robots.txt, the text file at the root of a site that tells automated visitors which parts they may read. GPTBot, PerplexityBot, Google-Extended and similar are separate from Googlebot, and blocking them removes the business from those assistants entirely. Many website platforms and hosts set these defaults without telling the owner. Suppose a fourteen-room hotel near Salcombe moves to a new site build one weekend. Its Google rankings can be unchanged on Monday morning while it has quietly dropped out of assistant answers, because a default in a configuration file decided the matter. This is a publishing decision, not a technical safety measure, and it is worth making deliberately rather than by accident.

Evidence density. This means whether pages carry attributable facts, figures and sources. The first controlled experiment on this, run against a 10,000-query benchmark, found that citing sources, adding quotations and adding statistics each raised visibility by 30 to 40% on position-adjusted word count (Aggarwal and others, ACM SIGKDD, August 2024). The 40% is a maximum, not an average. The same study found keyword stuffing scored roughly 8% below doing nothing at all, which is a long-standing habit actively working against the business paying for it. A later preprint held the wording identical and changed only structure, hierarchy and chunking, and reported citation rates rising 17.3% across six generative engines (Yu and others, 31 March 2026). That paper has not been peer reviewed and does not name the six engines, so treat it as directional. In practice, this is the difference between a dental page promising excellent results at affordable prices and one that states the material used, the number of appointments and the price.

Citation is also dispersed rather than concentrated. Reddit was the most-cited domain in Google AI Overviews and appeared in only 2.2% of responses, against 6.6% for Perplexity and 1.8% for ChatGPT (Ofcom, Online Nation 2025). No single site owns the answers, which is precisely why an ordinary business page carrying attributable facts has a route in.

It is worth saying what this work is not. For Google Search, Google lists things site owners can ignore: llms.txt and other special machine-readable files, chopping content into small pieces, rewriting pages specifically for AI, chasing inauthentic mentions, and over-focusing on structured data (Google Search Central). There is no special schema that makes a page eligible for AI Overviews or AI Mode.

How do you measure whether an assistant recommends you?

By asking it, repeatedly, and writing down what it says. The method is a fixed set of queries, run monthly, with each query asked five times within the run, across the major assistants rather than one. Answers vary between models, and they vary between runs of the same question on the same model, so a single answer is an anecdote rather than a measurement.

Ask “best commercial property solicitor in Exeter” five times in one week and you may get five overlapping but different sets of names. One appearance is not visibility and one absence is not failure. Repetition is what turns a set of anecdotes into something you can compare month to month, and it is also what tells you whether you are named consistently or occasionally.

What gets recorded matters as much as how often. Not only whether the business was named, but how it was described, which page was cited, and whether the description was accurate.

Checking one assistant is not the same as checking the field. In the year to August 2025, assistants grew from smaller bases at very different rates: Gemini up 146%, Claude up 138% and Perplexity up 100% (Ofcom, Online Nation 2025).

A conventional ranking report cannot substitute for this, for a simple reason: it records a position in a list of links. It has nothing to say about whether a name appeared in a written answer, and for ChatGPT or Perplexity there is no ranking position to report in the first place. Traffic reporting understates the picture too: a customer who reads an answer, forms a view and acts on it later may leave no trace at all in an analytics account.

The firstmention.co.uk five-model entity test

First Mention ran the test on its own domain. Five AI models were asked about firstmention.co.uk. Three of the five resolved it to firstmention.co, an unrelated US home-services SEO firm, and one attributed that company's content to First Mention.

This is worth sitting with, because it is an ordinary failure rather than an exotic one. The domain is spelled correctly. The site is indexed. Nothing is broken in any sense a conventional audit would flag. The models simply had a near-identical name in front of them, more material attached to the other one, and no unambiguous signal about which business sat in Devon and which sat in the United States.

It is also invisible by design. No ranking report contains a row for “confused with a company on another continent”, and no traffic report shows the enquiries that went elsewhere because the answer described the wrong firm. The only way to find it was to ask the models directly and read what came back.

The fix is identifiable work rather than a trick: the business name written in full and identically everywhere it appears, the country and county stated plainly on the page rather than inferred from a phone number, organisation markup generated from that visible text, verified external profiles that agree with each other, and enough distinct first-hand material that a model has something specific to attach to the name.

Any business in the south west of England with a common name faces the same arithmetic. A Riverside Dental, a Castle Garage or a Harbour Inn is competing for its own identity before it competes for anything else.

Questions people actually ask

More answers are on the First Mention FAQ hub.

← Back to all guides