Let’s talk

E-commerce · GEO

How to get into an AI’s recommendations: a GEO audit checklist for e-commerce

Shoppers increasingly choose not in Google or Yandex but in an answer from ChatGPT, Alice or Perplexity. Here is what happens to a store the AI simply doesn’t see — and what an e-commerce team can do to get its products into the recommendation.

When the machine does the buying

The pattern is simple. Instead of typing a query into a search engine, the shopper asks an AI: “Which robot vacuum should I get for carpet, under 30,000 rubles?” The decision is made right in the chat, often without a visit to any website.

Search has always split the world into “the top three” and “line forty”. An AI answer has, in effect, a single place: the recommendation in the text. Whether your product or a competitor’s gets there depends on how well the model reads your site.

This is GEO, generative engine optimization. It doesn’t replace SEO, it adds to it: SEO works for a position in the search engine, GEO for a presence in a language model’s answer. A study co-authored at Princeton University (the paper “GEO: Generative Engine Optimization”) showed that the right writing techniques raise how often a given product is mentioned in answers by up to 40%.

We are a digital agency focused on e-commerce: online stores, performance marketing, brand promotion. For the last few months almost every brief has carried an unspoken question: “are we in the AIs?” We checked how models read a catalogue and put together this checklist. It doesn’t depend on the niche: it works for a fashion store and for a marketplace alike.

How a model “reads” a store

First, what a model does not do. It doesn’t walk into your catalogue and scroll through 300 pages. It builds the answer from two sources: its own memory (training) and live data, when the service can follow links, as Perplexity or ChatGPT can.

For a “which one should I choose” question, the model relies on:

  • product pages: name, description, specifications, price;
  • reviews, as outside confirmation of the seller’s claims;
  • guides and comparisons, both your own and third-party ones;
  • mentions in independent sources: media, aggregators, marketplaces.

Models live in different environments. Alice and YandexGPT grew up inside the Yandex ecosystem and rely on it first of all; ChatGPT and Perplexity rely on Western sources. So check your visibility in both: a model that recommends you in Perplexity may stay silent in Alice’s answer, and the other way round.

Then the model cross-checks the data. If your site gives a drill a torque of 40 Nm, the marketplace says 28 Nm and reviews say “drives screws poorly”, the candidate looks unreliable, and the recommendation goes to someone whose data agrees. Hence the first rule of GEO in e-commerce: product data must not contradict itself across sources.

The checklist

1. The product page is the answer. The first two sentences are what the model is most likely to take into context when it assembles an answer. Write them so that a person could choose the product without opening the specification table: what it is, who it is for, how it differs. “Super quality” and “best price” don’t work: to a model they are noise, to a person they mean nothing. Before: “powerful, quiet, best in its class”. After: “1,200 W, 1.8 l tank, HEPA filter, 52 dB, 72 minutes on one charge”, “fits a 13-inch laptop backpack”, “IP67”. The first is only adjectives; the second is data that can be checked against other sources.

2. Structured data is up to date. Check your schema.org/Product and schema.org/Offer markup: name, description, price, availability, reviews. Running it through a validator takes 20 minutes. If the price in the markup is more than two days old, or the item is marked “in stock” while the warehouse is empty, to a model it is a stale product.

3. Reviews are a source, not decoration. “Great, buying again” helps neither the model nor a person. A review with a specification, a comparison with a competitor or a use case is data the model can lean on in an answer. Ask for specifics: in the post-purchase email, ask one question — “what did you buy it for?”

4. Guides and comparisons. If your niche has a “which one should I choose” question, there should be a page that answers it: selection criteria, a comparison table, a recommendation with reasons. According to the GEO study (Princeton and co-authors), structured, information-dense content like this is mentioned in AI answers noticeably more often than scattered paragraphs. On your site, such pages also work as a bridge into the catalogue.

5. Freshness in the text. Not in meta tags but in the text itself: “prices and specifications current as of August 2026”, in the first paragraph of a product page or guide. When a model picks a source, it relies on the text, not the metadata.

6. Mentions agree. The product name, key specifications and price range are the same on your site, on marketplaces, in aggregators and in the media. Contradictions lower the model’s confidence: it picks a “clean” candidate.

7. You have decided what crawlers see. Look at robots.txt: whom you block (GPTBot, PerplexityBot, Google-Extended) and why. Blocking “just in case”, without understanding, means losing out: your product won’t make it into answers, while a competitor’s product that the model can see will. Close confidential sections (wholesale prices, a supplier area) one by one, not the whole site.

8. llms.txt. A proposed standard: a short Markdown file at the root of the site that explains what the project is and in which sections the useful information lives. For a small store it’s half a day of work, for a large catalogue a proper task — but the model is spared a blind search and gets your catalogue’s structure.

9. Measurement. Make a list of 20–30 questions shoppers really ask: “which drill should I buy for home use”, “a light 27-inch monitor”. Once a month, run “blind” queries (without naming the brand) in ChatGPT, Alice and Perplexity and record whether your product makes the recommendations, whether there is a link to your site, and how often. A table works best: date; model; query; brand named or not; product included or not; link or not; which competitors are in the answer. Answers change from day to day, so a single screenshot proves nothing; what matters is the frequency over two or three months. That is the working metric: your share of mentions in AI answers, your GEO share of voice.

Five common mistakes

“We have SEO, it will cover everything.” SEO protects a position in the search engine, GEO a mention in the answer. They are different channels with different metrics.

Waiting for a contractor before sorting out your own data. Consistent data is the entry threshold; after that, clarity wins — the model picks whoever names the use case, the specifications and the difference from competitors precisely. The basis of your GEO is your product pages, reviews and prices.

Blocking crawlers “just in case”. That is a strategy by default, not by choice. At the very least, audit who reads your site and how often.

A one-off audit. Answers vary, and so does the market. Optimization works as a regular routine, and the metric is tracked over time.

Writing “for machine readability” everywhere. A person reads the product page first, and the model reads the same text. If a sentence is awkward for a person, the model doesn’t need it either.

What to do this week

  1. Your five best products: rewrite the first two sentences of their descriptions and complete the specifications.
  2. Run your schema.org/Product and schema.org/Offer markup through a validator and check prices and availability.
  3. Top guides: add an explicit “current as of…” date.
  4. Twenty shopper questions: a first blind test in ChatGPT, Alice and Perplexity.
  5. Look into robots.txt: who is blocked, who isn’t, and why.

The real goal

The goal is not a link to your site in an AI answer, although that is a nice bonus. The goal is the recommendation: a person who no longer googles but asks sees your product first on the list.

The niche is still open. Almost all recent GEO publications — the Sostav feed, the vc.ru tops, agency articles — are about B2B: logistics, services, PR. E-commerce appears in them as an example, not as the subject of a method. For a store whose product pages are its main asset, this is open ground. And the foundation of the work is what a healthy store has had for a long time: tidy product pages, real reviews, a catalogue with current prices. Whoever puts things in order now will be in AI answers before everyone else when AI recommendations become an ordinary shopping channel. Whoever waits until GEO is as standard as SEO will pay more to catch up.

Boson Group is a digital agency: online stores, performance marketing, branding, AI agents. Need a GEO audit of your product pages with a plan of recommendations? Write to team@boson.group.

More about Performance marketing at Boson Group

More from the blog

Contact

Send a brief — we reply with a time and cost estimate within 24 hours.

Brief

Tell us about the project.

Send a brief — we reply with a time and cost estimate within 24 hours.

What do you need?

* required