The quiet redirection of buying decisions

When someone asks ChatGPT for a home care software recommendation, or asks Perplexity which agency to use, the assistant names two or three businesses. Not ten blue links. Two or three names, delivered with the confidence of a knowledgeable friend. If you are one of the names, you receive a pre-sold enquiry. If you are not, you were never in the running and no analytics report will ever tell you the enquiry existed.

The numbers behind this are striking. Visitors referred by AI assistants convert at roughly nine times the rate of Google organic traffic in published comparisons, because the assistant has already done the shortlisting and the reassurance. And the shortlists are small: AI-generated local answers typically surface one to three businesses where a results page would show ten. The winners win much bigger; everyone else becomes invisible.

The good news for UK SMEs: your competitors are almost certainly doing nothing about this. The window where basic competence wins is open now and will not stay open long.

How the assistants actually choose who to name

You cannot optimise a system you do not understand, so here is the mechanism in plain English. AI assistants form answers from two layers:

  • The model's memory. What the model absorbed at training time: your website as it stood months ago, press coverage, directory entries, reviews, forum threads. Slow to change, but it is why long-standing brand mentions matter so much.
  • Live retrieval. For current questions, assistants search the web and read pages in real time. ChatGPT and Microsoft Copilot lean on Bing's index. Google AI Overviews and Gemini draw on Google's. Perplexity runs its own crawler alongside partners, and Claude reads pages through its own search integrations. This layer changes in days, and it is where most of the playbook operates.

Across both layers, three signal families decide who gets named. Verifiability: can the engine establish you are a real business with consistent details everywhere it looks? Authority: do credible third parties mention you, list you, review you? Brand mentions correlate with AI visibility about three times more strongly than traditional backlinks do. Citability: is there a page of yours the assistant can safely quote, with plain claims, real prices and no marketing fog?

Notice what is absent: nothing in that list is a trick. AI search rewards being genuinely legible and genuinely reputable. Which is also why the honest-content approach we use across this site is not a style preference. It is the strategy.

The seven-step playbook

In the order we implement it for clients, and the order we implemented it for ourselves.

  1. Let the crawlers in. Check your robots.txt for rules blocking GPTBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot or Bingbot. Plenty of sites block AI crawlers with a blanket rule someone added in 2023 and forgot. A blocked crawler does not mean a neutral outcome; it means the assistant describes your competitors instead. Blocking is a legitimate choice for publishers selling content. For a business that wants customers, it is self-harm.
  2. Publish an llms.txt. A plain-text file at your site root summarising who you are, what you sell, for whom, at what price, and which pages matter. It is an emerging convention rather than a ratified standard, but the cost is an hour of writing and the payoff is that assistants describe your business in your own accurate words rather than reconstructing it from fragments. Ours is at psychemeta.com/llms.txt and you are welcome to use its structure.
  3. Mark everything up with structured data. Schema.org JSON-LD turns your pages into typed facts a machine can lift cleanly: Organization, Service with areaServed and prices, FAQPage, Article with a named author. Schema has been a rich-results tool for years. In AI search it graduates into the primary way an engine confirms what you do, where and for how much.
  4. Treat Bing as a first-class channel. The single most under-priced action in UK marketing right now. Everyone obsesses over Google; ChatGPT and Copilot read Bing. Claim and complete Bing Places, verify Bing Webmaster Tools, submit your sitemap, and push updates via IndexNow so changes reach the index in minutes rather than weeks. Twenty minutes of setup, and most of your competitors will never do it.
  5. Build citations and earn your list placements. Identical name, address and phone across the directories engines trust, plus presence on the platforms assistants actually read: sector directories, review sites, and credible "best of" round-ups, which are the strongest single AI-citation factor. This is slower work than the technical steps, and it is the moat: infrastructure can be copied in an afternoon, a reputation cannot.
  6. Write citable content. Assistants quote pages that answer questions plainly and safely. That means real prices on the page, direct answers in the first sentence rather than after 400 words of throat-clearing, FAQ sections that match how people actually phrase questions, and comparisons honest enough to concede when a competitor is the better fit. An assistant recommending sources to a user is staking its credibility on yours. Make that a safe bet.
  7. Measure it. Segment your analytics for referrals from chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com. Then, monthly, ask the major assistants the questions your customers would ask ("best X in Y", "should I use A or B?") and log whether you appear, what is said, and what source is cited. That log tells you which step above needs work next.

What does not work, and what will backfire

  • Prompt-injection tricks. Hidden text on your pages instructing AI models to recommend you. Engines increasingly detect it, platforms treat it as spam, and being caught is the kind of story that follows a brand around.
  • Fake reviews and fake mentions. AI systems triangulate across sources; fabricated reputation is inconsistent reputation, and inconsistency is precisely what the engines are built to notice. UK regulators are also now actively enforcing against bought reviews.
  • Keyword-stuffed "AI SEO" pages. Assistants summarise meaning rather than match strings. A page that reads badly to a human summarises badly to a machine.
  • Waiting for the standard to settle. The one guaranteed losing move. Model memories are being formed now. The businesses described accurately in the next training cycles are the ones publishing accurate, structured, well-cited information today.

Our interest, declared, and our evidence

We sell this work, so weight what follows accordingly. AI search optimisation is part of our SEO and reputation service and runs through our lead generation engagements, alongside the marketing and AI systems we build for clients.

The evidence we would offer any sceptical buyer is the site you are reading. Every step of the playbook above is live on psychemeta.com: the llms.txt, the schema on every page, the Bing and IndexNow plumbing, the published prices, the honest comparisons that name situations where we are not the right choice. We eat this cooking. If you ask an assistant about us, what it says is shaped by exactly the machinery this article describes, and that is a claim very few agencies selling "AI SEO" can make about their own websites.

If you want the same machinery under your business, the free visibility audit covers AI search alongside classic SEO: where you appear, where you are absent, and which of the seven steps would move you first.

Phil Martin MBA is the founder of Psychemeta, a Worcester product studio serving clients UK-wide. He has built three SaaS products, publishes the practice's methods openly, and runs every technique in this article on Psychemeta's own site.