Do search engines and AI assistants understand your store?

Since May 2025 Google has shown AI Overviews in Czech above search results, and since October 2025 it has offered a separate AI Mode. Store owners are therefore being offered “AI optimisation”: new markup, special files, rewritten copy. Google, however, states explicitly that there are no additional requirements for appearing in its AI features. A machine needs the same thing as a shopper: a page that answers their real questions and facts that never contradict each other. Where that is missing, markup will not help. Where it is present, markup simply makes it easier to read.

What has actually changed in search

AI Overviews have been available since 20 May 2025 in more than 200 countries and more than 40 languages, including Czech. Google launched AI Mode in Czechia on 7 October 2025; according to Google, people ask questions two to three times longer than in classic search, and answers include links to sources. A longer query such as “a quiet coffee grinder under CZK 3,000 that is easy to clean” is exactly where it matters whether a page carries specific facts or only a generic description.

How much this changes traffic is not clear-cut. Pew Research Center tracked the browsing of 900 US adults in March 2025: when an AI summary appeared above the results, people clicked a traditional result in 8% of visits, compared with 15% without one. They clicked a link inside the summary in 1% of visits. Google, by contrast, says clicks from results pages with AI Overviews are higher quality and that people spend more time on the site. A US sample and general queries cannot be transferred directly to a Czech store. One conclusion holds, though: a citation in an answer is visibility, not a visit.

Measurement only helps so far. Visits from Google’s AI features are counted together with other search results in Search Console, and what an answer says about a store when nobody clicks cannot be measured by anyone.

Machines need what shoppers need

For AI features, Google recommends the same as for search in general: unique content that answers people, a clear page, access for Googlebot and important information in text rather than only in images. Microsoft adds, for Bing and Copilot answers, that AI assistants do not read a page top to bottom but break it into smaller pieces. It recommends clear headings, lists and tables and warns against hiding key information in tabs, PDFs or images.

In practice this means finding information gaps. An illustrative, invented example: a coffee grinder page has a name, price, photos and two paragraphs about “the perfect experience”. In searches and support tickets, customers ask about noise, grind settings, hopper capacity and cleaning. The noise level appears only on a photo of the box, and the cleaning guide sits in a PDF. A person cannot find the answer, an assistant cannot read it, and the answer comes from another website.

Where to find what people ask: on-site search queries, customer support questions, return reasons and reviews. The answers then belong in the page text, and specifications in the specification table. Fix the page for the person first; only then deal with how it reads to machines.

Facts must match everywhere

Several sources now describe the same product: the page text, the structured data in its code, the feed for Merchant Center and comparison sites, and the terms and conditions. Google requires structured data to match the visible content of the page; misleading markup can lead to a manual action and the loss of rich results. For pages where a product can be bought, Google lists name, image and an offer with price and currency as required, and recommends shipping, return policy, GTIN and brand. Combining structured data with a Merchant Center feed, it says, maximises eligibility for shopping experiences.

Seznam, the Czech search engine, also reads product details (image, price, rating, availability) from schema.org markup in the source code and recommends that a page carry data for only one matching product. The schema.org product vocabulary itself includes identifiers such as GTIN and SKU, brand and additional properties.

Shoptet has stated since 2013 that its templates mark up basic product details automatically. Owners therefore usually do not need to write markup, but they should check it: template customisations or add-ons can break or duplicate it. You can check a product page’s code in Google’s Rich Results Test.

The most common contradictions involve the sale price, availability, shipping cost and return period. If the page says “in stock”, the feed says “within a week” and the terms state a different return period from the page, every machine gets a different answer. Fix the source in the store, not each individual output.

Content beyond products: who, when and why

For guides, comparisons and the blog, Google recommends making it clear who wrote the content and, if automation helped, how and why. It also warns against changing a page’s date to make it look fresh when the content has not substantially changed. State the author or editorial team, the publication date and genuine revision dates, and for comparisons, what they are based on. This is about reader trust, not a technical trick.

Crawlers: search and training are two decisions

A robots.txt file tells crawlers where they may go. For AI it separates two questions. OpenAI uses OAI-SearchBot to surface websites in ChatGPT search; a site that blocks it will not appear in ChatGPT search answers. GPTBot, by contrast, collects content for model training, and blocking it does not decide search visibility. Similarly at Google, the Google-Extended token for Gemini and Vertex AI does not affect inclusion in Google Search.

Check that neither robots.txt nor bot protection at your host or CDN excludes crawlers you want. And remember the limits: robots.txt is not access control, and not every crawler follows it. OpenAI also notes that visits triggered directly by a user’s request may not follow robots.txt rules.

What markup will not do

Google does not guarantee that structured data will be shown, even on a correctly marked-up page. A telling example: since August 2023 FAQ rich results have appeared only for well-known, authoritative government and health websites. For AI features, Google writes explicitly that you do not need new machine-readable files, AI text files or special schema.

Microsoft is more positive about markup and recommends JSON-LD schema because it helps AI systems understand what type of content a page holds. It too adds that nothing guarantees selection. A sensible conclusion from both views: markup is a translation of the page into machine language. If the page has no answer, the translation will not create one.

What to do now

1. Pick ten important pages. Products and categories with the most demand or revenue.
2. Find the information gaps. For each page, list the three most common customer questions and check whether the text answers them.
3. Compare the facts. Price, availability, shipping and returns on the page, in the feed and in the terms; fix any contradiction in the store.
4. Check markup and crawlers. Run the pages through the Rich Results Test and check that robots.txt decides separately on search and on training.
5. Measure outcomes, not promises. Track organic visits and orders from these pages and visits from AI assistants, not the number of tags.

In its SEO & GEO section, Korzaro shows which tracked queries trigger a Google AI Overview that cites the store and in which position, alongside visits, orders and revenue from AI assistants according to GA4. From the public robots.txt it also shows whether the store has taken any position on AI crawlers. It cannot see ChatGPT or other assistant answers that nobody clicks – a measurement limit no tool can cross.