More and more people no longer search. They ask. “Who is a good accountant near me?”, “which web designer should I use for a small business?”, “recommend a roofer in Chelmsford.” ChatGPT, Perplexity, Gemini and Google’s AI Mode answer with a shortlist of businesses by name, and most people act on it without opening ten blue links.

Zero-click searches, queries that end without a user clicking through to any website, now represent 68% of all Google searches, according to Similarweb’s latest data.

For a small business, that shortlist is the new front page. It is also the least understood surface in marketing, which is why so much of what is written about it is either vague or a sales pitch. This guide is the mechanism: how the assistants actually decide which businesses to name, why the lists barely overlap, and the specific, testable things a UK small business can change.

There is no AI ranking, and that is the whole point

The first thing to unlearn is the idea that there is a single algorithm to beat. There is not. Each assistant assembles an answer at the moment you ask: it searches the live web, pulls a handful of sources it can reach, and then a model decides which few are worth naming. The sources and the reasoning differ by product, by question and by day.

You can see this in the raw behaviour. In my own baseline — fifteen questions put to four surfaces, sixty answers — Perplexity cited eighteen to twenty sources on almost every answer, while Gemini returned no citations at all on seven of fifteen despite grounding, and Google AI Mode returned no references on six.1 One product shows its work and cites heavily; another names sources in prose and cites none. There is no shared scoreboard, so there is no shared ranking to climb.

What sits underneath all of them, though, is consistent. Before an assistant can name your business, three things have to be true.

The three things an assistant needs before it can name you

  1. It has to find you. Somewhere on the open web there must be a page that connects your business to the question — your own site, a directory, a profile, a mention. If nothing reachable says “this business does this, here”, you cannot appear in the answer, however good you are.
  2. It has to understand you. A machine has to be able to extract what you do, where you do it and what you are called. Consistent names, addresses and phone numbers; a clear list of services; structured data that states the facts in a parseable form. Ambiguity is invisible.
  3. It has to prefer you. When several businesses fit, the model chooses. What tips it is corroboration and specificity: several independent sources that mention you, and a page that answers the question directly rather than gesturing at it.

The encouraging part is the level of the bar. In my baseline the most-cited domain across all sixty answers appeared six times. Webflow, Semrush, HubSpot, YouTube and Reddit were among the leaders, and small UK agency blogs were cited right alongside them.2 Nobody dominated, and authority was not the deciding factor. A specific, well-framed page that answers the question plainly beat size.

What the assistants actually pull from

When you pull apart the answers for a local or niche question, the sources fall into a few recognisable groups:

  • Your own pages — but only the ones that answer the question in plain, readable text, not behind a form, an image or a script.
  • Profiles and directories — Google Business Profile, trade bodies, industry directories, review sites. Google AI Mode answered the Essex agency question I tested with twenty Maps listings; the language models leaned on directory-style pages.
  • Third-party mentions — “best X in Y” listicles, local press, comparison posts, supplier and partner pages. A single mention in a credible list can do more than a year of blog posts.
  • Structured data — the schema on your pages that states your name, services, location, opening hours and reviews in a form the machine does not have to guess at.
  • Public discussion — Reddit, forums and Q&A sites. Reddit was among the four most-cited domains in my baseline, and it is not a channel most small businesses think of as marketing.

Notice how much of that is not your website. This is the part that gets missed when “AI SEO” is sold as a content programme. Being named by an assistant is at least as much an off-site job — being legible and mentioned in the places the assistants already trust — as an on-site one.

Why the recommendation lists barely overlap

The clearest evidence that there is room for a small business is how little the answers agree with each other. Across four “recommend an agency” questions, the same few names did not recur. The SEO-audit question produced twenty-two different agency names across the four surfaces. The research-led web design question produced fifteen, none of them twice. The AI-SEO question produced more than thirty, with a handful repeating.

Then I compared that with Google. Of the eight agencies ranking on page one of Google for the same commercial terms, only three appeared anywhere in the AI answers — and one of those was named but never cited. The set of businesses an assistant recommends is largely a different set from the one Google ranks.3.

The practical reading: you are usually not trying to displace a well-entrenched incumbent. You are trying to be legible enough to be assembled into an answer that currently names a near-random handful of neighbours. That is a much more winnable position than “rank number one”.

What a small business can actually change

Strip away the jargon and the work comes down to six moves. None of them is a trick, and all of them are checkable.

1. Answer the exact question on a page of your own

Write the page a buyer’s question deserves. Put the answer in the first paragraph, in text, in the words they used. One question, one page — cost, comparison, “do I need one”, checklist. The pages that get cited are almost always the ones that answer directly rather than the ones that describe how wonderful the company is. My own research found that the question “what research should I do before redesigning my website?” had no UK small-business answer at all; the field was American SaaS guides. That absence is an opportunity.

2. Make the business machine-legible

Decide on one exact business name, address and phone number and use them identically everywhere. Complete your Google Business Profile properly. Give each service its own page. Add structured data — Organization or LocalBusiness, Service, FAQPage — so the facts are stated rather than inferred. If a machine cannot tell what you do from your homepage in one pass, neither can the answer.

3. Be mentioned where the assistants already look

Directories, trade associations, local press, supplier pages, genuine review sites, and the “best X in Y” lists your buyers actually read. A mention in a page the model already cites is worth more than a new post on a site it has never seen. This is slow, unglamorous work, and it is most of the job.

4. Keep it current

Assistants favour fresh signals. Date your pages, review them, and update the numbers. A pricing guide from 2022 that says “from £500” is less useful than one reviewed this quarter, and it reads that way to a model as much as to a person. A visible last-reviewed date and a genuine update cycle are cheap and they compound.

5. Let the crawlers reach you

Check that your robots settings do not accidentally block the bots assistants use, and that your content is server-rendered. Many AI crawlers do not run JavaScript, so text that only appears after a script loads may as well not exist. This is one reason this site renders its content on the server: so the machines read the same words you do.

6. Measure it, because it is measurable

Write down fifteen questions your buyers would ask, put each to three or four assistants, and record every business named and every page cited. Count them. Do it before you change anything and again each quarter. It is the cleanest before-and-after measure that exists for this work, and almost nobody bothers to take it.4.

What does not work

  • Writing more blog posts about yourself. Volume without a specific answer is noise.
  • Keyword-stuffing pages meant for machines. The models read the page; they can tell.
  • Buying citations or placements that no human trusts. Assistants draw on sources that are corroborated, not bought.
  • Assuming Google rankings carry over. In my research the AI-recommended set barely overlapped with the Google set. A position one ranking is not a citation.
  • Treating it as a one-off. The surfaces change month to month; whatever earned a citation in March is not guaranteed in September.

The businesses that lose here are usually not doing anything wrong; they are simply invisible in machine terms. No schema, no directory, no page that answers the question. The assistant has nothing to work with, so it names someone else.

How this relates to the rest of the work

Being recommended by an assistant is the visible end of a longer chain. The assistant can only cite facts that are on your site and in the sources it trusts, which means the foundations are the same ones that make a site work in ordinary search: know what your market asks, put the answer where it can be found, and keep it true.

That is why I treat it as three stages. GroundWork maps what your market actually searches for and who the assistants already cite for your questions. Build puts the pages and the structured data in place, server-rendered and owned by you. Coverage is the ongoing part: keeping the pages current, the mentions flowing, and the citation count moving — measured against the same question set, quarter after quarter.

If you want to see the shape of the starting point, there is a full sample Demand Map on this site, written for a development consultancy and anonymised. It is the document that says which questions are worth answering before anyone writes a page.

The short version

  1. Be findable: a page somewhere reachable that ties your business to the question.
  2. Be understandable: one name, one set of facts, stated in structured data.
  3. Be corroborated: mentioned in the sources the assistants already trust.
  4. Be current: dated, reviewed, updated.
  5. Be measurable: the same question set, counted every quarter.

None of this is fast, and none of it is mysterious. It is the same discipline as always — be the clearest answer to a real question — applied to a search surface that no longer returns ten links.

The search surface changed. The work did not.

Where to go next

  • Coverage — the ongoing AI SEO work: legibility, mentions, freshness and a measured citation count.
  • GroundWork — the research that says which questions your market asks and who the assistants cite for them: £250, one Demand Map.
  • Build — the site that states your facts clearly, server-rendered, with the structured data the machines need.
  • Rented Ground — what you actually own when you own a website, and why that matters for something as long-lived as search.

The supporting guides in this series — generative engine optimisation, Google AI Overviews, what SEO costs now, and whether SEO is dead — will be linked here as each one publishes.

  1. KOODOS AI-visibility baseline, 17 September 2026: 15 questions × Perplexity, ChatGPT, Gemini and Google AI Mode, 60 completed calls. Raw responses and a per-call citation index in docs/research/ai-visibility-2026-09/. ↩
  2. KOODOS AI-visibility baseline, 17 September 2026. Top cited domains: webflow.com 6/60, then Semrush, HubSpot, YouTube and Reddit at 4. ↩
  3. KOODOS competitor gap and AI-visibility baseline, September 2026. AI-answer name sets per question from the 60-call baseline; Google page-one set from the twelve-domain DataForSEO pass. ↩
  4. The method and raw data are published in the KOODOS AI-visibility baseline, 17 September 2026. ↩