What gets a local business named by an AI assistant

Ask an AI assistant to recommend an accountant in Gloucester and it will name three or four firms. There are far more than four accountants in Gloucester. Something decides which ones get said out loud, and most advice about it is guesswork dressed as strategy.

We are in an unusual position to check, because we already measure six other things about every business in the Local Digital Visibility Index. So we can take the businesses an assistant names, take the ones it never mentions, and ask what actually differs between them.

Across 1371 businesses in 34 South West indices, 355 were named in at least one answer and 1016 were never named at all. That is 26% — the assistant recommends roughly one business in 4.

Every figure on this page is computed from the published datasets when the page is built, so it tracks the index rather than describing an earlier version of it.

The finding

Median pillar score for the 355 businesses named by at least one AI assistant, against the 1016 never named. The line is the gap. Speed is the pillar that barely moves.

Named by an AI assistant Never named

Pillar	Named	Never named	Gap
Speed & CWV	61	64	-3
Technical	88	75	+13
Local presence	83	73	+10
Visibility	37	0	+37
Content & trust	55	41	+14

The named businesses score 64 overall against 49 for the unnamed ones. That much is unsurprising: better businesses online get recommended more. The interesting part is which pillar carries the difference and which does not.

Speed & CWV moves by -3 points. On a 0–100 scale, across 1371 businesses, that is nothing. The firms an assistant recommends load at almost exactly the same speed as the firms it ignores.

Visibility moves by 37 points, which is the largest gap of any pillar we measure.

If you have been told that making your site faster will get you recommended by AI, this is the check on that claim, and it does not support it.

Why speed still matters, just not for this

The honest reading of a two-point gap is narrow: within the range these businesses actually occupy, speed does not distinguish the recommended from the ignored. It is not a finding that speed is worthless.

Speed still decides whether someone waits for your page after they arrive, and that is a conversion question rather than a discovery one. We have written up what page speed does and does not do separately, and the same caution applies there. The mistake is treating a real usability property as if it were a ranking or recommendation lever.

There is also a range problem worth naming. Almost every site in this cohort is somewhere between mediocre and decent on speed; hardly any are catastrophic. A test that includes no catastrophic sites cannot tell you whether being catastrophically slow would hurt. It almost certainly would.

Ranking and being named are not the same channel

The obvious objection is that this is one finding wearing two hats — that assistants simply repeat whatever ranks, so the visibility gap explains everything and there is nothing else to see.

The data says the two overlap heavily, but not symmetrically, and the asymmetry is the useful part:

• 411 businesses rank organically and are never named. They are findable the traditional way and still absent from the answer.
• 31 businesses are named while ranking for nothing. Twenty-six, out of 1371.

Read those together and the shape is clear. Ranking is close to a precondition for being named — almost nothing gets recommended without it, and 31 exceptions in 1371 businesses is thin enough that it could be measurement noise rather than a route anyone could take deliberately. But ranking is nowhere near sufficient: 411 businesses have cleared that bar and are still never mentioned.

So the practical point is not the comfortable one that these are two independent channels you might win separately. It is narrower and more demanding: being on page one buys you a ticket to be considered, and most businesses holding that ticket are still not chosen. Whatever decides the second step, it is not the thing that got them through the first.

What this does not prove

This is ten local sectors in the South West, measured once, and it is correlational throughout. Nobody made a site faster to see what happened.

Three specific limits, stated because they are the ones that would change the conclusion:

1. One assistant. Every figure here comes from Perplexity (model sonar) on five prompts per index. It is not evidence about Gemini, ChatGPT or AI Overviews, and our methodology page was corrected on 30 August 2026 for having implied otherwise. A business absent here may be named elsewhere.
2. The pillars are entangled. Content and Visibility both move a long way, and they are not independent of each other — a firm with real pages about what it does tends to rank and to be quotable. This data cannot tell you which of the two is doing the work, or whether a third thing causes both.
3. Local presence carries a sector effect. The Local pillar includes review velocity, which was miscounted for most of this quarter and has since been corrected. The corrected figures are sound, but that component is substantially a property of the trade — restaurants and estate agents collect Google reviews continuously, builders and accountants largely do not — so the Local row above compares businesses from sectors with very different review cultures.

Five prompts is also a small basket. Ask a differently-worded question and a different set of names comes back; that instability is a genuine property of the thing being measured, not noise we have failed to remove.

What actually seems to move it

Given the caveats, the defensible reading is that assistants name businesses that have something to say and somewhere to say it — the two pillars that move most are the ones about published, checkable content and about appearing in results at all.

That is consistent with how these systems work: an answer engine assembles a response from sources it can find and attribute. A business with three pages and no substance offers nothing to attribute, however fast those three pages load.

Practically, and in order of how much the data supports it:

• Publish pages that answer the question a customer actually asks, in the words they ask it. Our note on structured data covers making those pages legible to machines once they exist.
• Make the basic facts about your firm explicit and on your own site — who you are, what you are accredited to do, where you work. Internal linking to about, team and credentials pages turns out to separate this cohort more sharply than speed does.
• Treat organic ranking and AI recommendation as two jobs that overlap, not one job with two names.

None of that is novel advice. What is new is that it is the advice this dataset supports, and the popular alternative — optimise the technical layer and wait — is the one it does not.

Where your own firm sits

All 1371 measured businesses are listed on the indices, each with its own scorecard.

If your sector or town is not in the index yet, the same six pillars can be run against it. The methodology is public and the pipeline is open source, so the figures above can be recomputed by anyone who wants to check them.