Published · Updated

How AI Search Engines Choose Businesses: What Retrieval Actually Rewards

No assistant publishes a business-ranking formula. Here is what AI retrieval actually rewards, shown on a real headless Shopify storefront, with the honest limits of visibility work.

By Tyler Stocks · Stocks Local

AI search engines do not publish one formula for choosing businesses. Each assistant retrieves sources from its own index, then generates an answer from the passages it trusts. In practice, retrieval rewards crawlable pages, direct answers to buyer questions, consistent business identity, accurate structured data, and independent corroboration. None of those factors guarantees a mention.

That is less exciting than a secret ranking-factor list. It is also more useful, because every part of it can be inspected and improved on your own site this week.

There is no universal AI business ranking

Google AI Overviews, Google AI Mode, ChatGPT search, Perplexity, Gemini, and Copilot are different products. They do not share one index, one retrieval system, or a permanent set of business-ranking weights. The same prompt can produce a different answer after a change of location, date, wording, or product mode. A provider reporting a guaranteed position without fixing those variables is not measuring a stable result.

What the platforms actually document is thinner than most GEO checklists suggest. Google states that its established SEO guidance still applies to AI Overviews and AI Mode, and that those features carry no additional technical requirements. That puts crawlability, useful content, internal links, and page experience ahead of novelty files and invented AI tactics. Read Google's guidance for AI features directly. It is a better baseline than a vendor promising a private submission route.

Other assistants publish crawler controls and product documentation, but none publishes a complete, permanent formula for which business gets named for a given prompt. I treat every universal checklist, including this one, as a working hypothesis rather than a specification.

What retrieval actually rewards

Assistants that cite businesses broadly work in two stages. First they retrieve candidate sources for the question. Then they generate an answer from the passages they trust. You cannot control the generation step. You can influence retrieval, and it rewards five things you can inspect on any website.

LayerWhat retrieval rewards
AccessPages that return a normal response, render useful HTML, and stay indexable
RelevancePassages that answer the buyer's question completely
IdentityA business that describes itself consistently everywhere it appears
LabellingStructured data that matches the visible page
CorroborationReviews, mentions, and links the business does not control

Access comes before optimisation

A useful page cannot be selected if the system cannot fetch it. Important product, service, comparison, and case-study pages should return a successful response, render their key facts in HTML, use a stable canonical, avoid accidental noindex directives, and receive internal links from relevant pages.

An llms.txt file can give a concise publisher-written summary to systems that choose to read it. It is not a submission endpoint and it cannot replace crawlable pages. And access alone creates nothing. It only makes evaluation possible. The page still has to answer the request better than competing sources.

Relevance means a complete answer to one question

A retrieval system looking for sources needs a page that actually answers the question asked. A service page should explain who it is for, what it includes, how it is delivered, how pricing works, what is excluded, what proof exists, and the next step. A product page should cover specifications, variants, materials, delivery, returns, warranty, and comparison.

Because these systems often work at paragraph level, each passage should make sense on its own, and each URL needs one distinct job. One strong page usually beats five near-duplicate articles chasing the same phrase.

Identity means describing the business the same way everywhere

The website and important external profiles should agree on the business name, offer, founder, location or service area, category, and canonical URL. We feel this one directly. Stocks Local can be confused with stock-market language, so clear titles, organisation markup, founder details, and matching profiles do real work in identifying the company. Consistency removes ambiguity. It does not create authority by itself.

Structured data labels facts, it does not create them

Structured data gives machines explicit labels for information that is already on the page. Google's structured data guidelines require markup to represent the visible content, warn against misleading information, and state that valid markup does not guarantee a rich result.

Mark up the organisation, product, article, breadcrumb, or FAQ only when the visible page carries the same information. Adding unsupported ratings, invented prices, or a keyword-heavy graph is not optimisation. It is a trust problem written in JSON.

Corroboration is the part you cannot manufacture

Every company can call itself an expert. Useful corroboration includes a live client project, a case study that names the problem and the decisions, specific reviews with a verifiable source, client or partner links, named authorship, and primary documentation behind technical claims.

For local discovery, Google says relevance, distance, and prominence shape local results, and that complete, truthful business information helps. Read Google's local ranking guidance before treating a profile as a shortcut. Avoid bulk directory packages, paid link schemes, and copied press releases. They create volume without corroboration.

A worked example: Awaken Saunas

Abstract advice is easy to nod along to, so here is how it looks on a real storefront. We built Awaken Saunas a custom Next.js storefront with Shopify behind the catalogue and cart, for a County Tyrone workshop selling handcrafted saunas across the UK and Ireland.

The 16-product catalogue is organised into outdoor, indoor, wood-fired, commercial, premium, and cold-water therapy paths, so a retrieval system, like a buyer, can reach the right category without wading through everything else. Product pages put price, dimensions, heat-up time, heater requirements, delivery, assembly, warranty, and custom sizing options in plain crawlable HTML rather than behind scripts. The build exposes canonical URLs and a sitemap, with business, product, breadcrumb, and FAQ structured data matching what each page visibly says.

None of that guarantees Awaken a mention in any assistant. What it does is remove every avoidable reason for a system to skip the site. The facts an answer engine needs about a handcrafted sauna are reachable, specific, and consistent. The full build is documented in the Awaken Saunas case study.

Claims that should make you suspicious

The gap between what platforms document and what some vendors promise is where budgets disappear. Do not accept:

  • guaranteed ChatGPT, Perplexity, or Google AI visibility;
  • a fixed timeframe for inclusion;
  • a private submission route into any assistant;
  • schema presented as a ranking switch;
  • the idea that opening every training crawler creates recommendations;
  • one prompt screenshot offered as proof of market visibility.

The credible promise is narrower: better access, clearer information, stronger evidence, and more useful measurement. That is the standard we hold our own GEO and AI visibility service to, and it is the standard to hold any provider to.

Measure referrals, not screenshots

Assistant answers vary by prompt, location, account, time, and product mode. A screenshot of one successful prompt is an observation, not a stable position.

If you monitor mentions, keep a fixed, repeatable prompt sample as a secondary signal. Ahead of it, track identifiable assistant referrals, organic landing pages, contact actions, qualified enquiries, proposals, and won work. A mention that produces no relevant visit is not a commercial result.

Where to start

  • Confirm priority pages can be crawled, rendered, indexed, and internally linked.
  • Give each URL one distinct question and answer it in the opening paragraph.
  • Make the offer, products, service area, founder, and contact details unambiguous.
  • Add structured data only where it matches visible content.
  • Publish real project evidence and cite primary sources for technical claims.
  • Earn relevant independent mentions and reviews.
  • Track referrals and qualified enquiries before claiming progress.

The businesses most likely to appear in AI answers are not the ones using the most GEO terminology. They are the ones with the clearest useful pages and the strongest verifiable evidence.

For the full method, read the complete guide to GEO. If you want a first review of your own store, request a free written Shopify teardown and see what an answer engine can actually read on your site.

Questions

Asked and answered.

  • Is it true that only 1 percent of businesses appear in ChatGPT?

    No public dataset supports a one-percent figure. It circulates as a marketing claim without a defensible source. What is observable is that assistants cite a narrow set of sources per answer, so most businesses do not appear for most prompts. The controllable factors are crawl access, clear answer-shaped pages, consistent entity details, and independent evidence.

  • Can anyone guarantee my business appears in AI answers?

    No. Assistant answers vary by prompt, location, account, and model version, and no supplier controls retrieval. Treat any guaranteed placement, guaranteed citation, or fixed AI ranking promise as a red flag. The honest work improves access, clarity, and evidence, then measures referrals and enquiries.

Want a clear second opinion on the site?

Get a free Shopify teardown