The most important shelf for your brand is neither in retail stores nor in your own shop. It is in an answer that ChatGPT or Google’s AI Overview puts together in a few seconds. In many companies, nobody knows whether their own brand appears in it.
Take a fictional manufacturer of kitchen knives as an example: its own shop, 400 products, twelve specialist retailers. When someone asks ChatGPT for a good chef’s knife up to €150, the answer names three brands, not this one, and cites two retailers and a review portal as sources. The knives are no worse. The portal lists the hardness ratings in a table, the manufacturer in a PDF.
So when it comes to being named in an AI answer, you are up against not only other brands but also Amazon, Otto and every retailer that stocks you. How unevenly power is distributed in online retail is shown by the bevh (German E-commerce and Distance Selling Trade Association) figures for 2025: manufacturers selling direct turned over €2.9 billion in German e-commerce, online marketplaces €46.2 billion (bevh, presentation at the 2026 annual press conference, slide 10). If you have opted for selling direct as a manufacturer, you know these opponents. But you own something none of them has: the truth about your product.
What is changing in search
0.94%. That was the annual average organic click-through rate for informational queries when Google showed an AI Overview and the brand was not cited in it. AI Overview is the original English name; in German search the feature appears as “Übersicht mit KI”. With a citation the rate was 2.07%, without an AI Overview 3.35%. The figures come from Seer Interactive’s 2026 update of April 2026: 5.47 million queries from 53 clients, January 2025 to February 2026 (Seer Interactive, AIO Impact on Google CTR, 2026 Update).
Seer itself qualifies the headline figure: a single account makes up 47% of impressions in this segment, and without it the value is around 1.6%. That does not change the ranking. Even 1.6% is clearly below the 2.07% with a citation and the 3.35% without an AI Overview.
The overall trend from the same analysis: for queries with an AI Overview, the organic click-through rate fell to a low of 1.31% by December 2025. It has been rising again since then, but at 2.36% in February 2026 it was still below the starting value of 3.19%.
Pew Research arrives at the same picture with a different method: in a US panel of 900 adults, 8% of visits with an AI summary in March 2025 included a click on a traditional result, compared with 15% without a summary (Pew Research Center, 22 July 2025).
Two caveats belong here. As early as its November 2025 update (data as of September), Seer pointed out that citation and click-through rate are only correlated; strong brands often bring both (Seer Interactive, September 2025 Update). And all values come from Seer clients or from a US panel, not from the German market. Anyone who turns this into “the top 3 capture 50%” has added up click-through rates that cannot be added up.
How AI systems choose their sources
What the systems officially require is surprisingly little. Google writes that there are “no additional requirements to appear in AI Overviews or AI Mode”: no special Schema.org types, no AI text files. A page has to be indexed and eligible to be shown with a snippet in search, and structured data has to match the visible text (Google Search Central, AI features and your website, as of 10 December 2025).
OpenAI separates its crawlers: OAI-SearchBot decides whether a page appears in ChatGPT search, GPTBot relates to training, and the two can be controlled separately in robots.txt (OpenAI, Overview of OpenAI crawlers). Whether your website is readable for these crawlers at all is covered in the article on AI visibility.
What the systems then actually select is documented by no one. The rest of this section is therefore my own assessment, not a statement from the providers. I see three levels.
First, reachability: the crawler is allowed in, the page renders server-side, and the content is in the HTML that is delivered.
Second, consistency of the entity. Back to the knife manufacturer: its PDF says 58 HRC, a retailer typed in 60 HRC when creating the listing, the marketplace says “very hard”, and the review portal measured 58 HRC. Four versions of one attribute.
No provider discloses which rule a language model applies to resolve contradictions like these. It seems plausible that the version that occurs most often wins, and that is often not the manufacturer’s but the retailer’s, whose record appears in dozens of shops. In the better case, the answer then names your knife with the retailer’s hardness rating; in the worse case, it names a different knife.
Third, citability: paragraphs that answer a question on their own, with a number, a unit and context. “High-quality steel for the highest demands” gives the model nothing it could cite. “The blade has 58 HRC: hard enough to hold its edge for a long time, and soft enough to be honed on a sharpening steel” does.
Product data and structured data: the GEO homework for D2C manufacturers
This is where GEO for manufacturers parts ways with what Generative Engine Optimisation means for media or service providers: your citable unit is the product. For products, Google and OpenAI have published concrete field lists, the merchant listing markup (Google Search Central, Merchant listing structured data) and the specification for the OpenAI product feed (OpenAI, Products feed spec).
| Field | Google merchant listing (markup) | OpenAI product feed |
|---|---|---|
| Name / title | Required (name) |
Required (title) |
| Image | Required (image) |
Required (image_url) |
| Price with currency | Required (offers.price, priceCurrency) |
Required (price, format “79.99 EUR”) |
| Availability | Recommended (availability) |
Required (availability) |
| Brand | Recommended (brand) |
Required (brand) |
| GTIN | Recommended (gtin) |
Optional (gtin), must have a valid check digit |
| Description | Recommended (description) |
Required (description, plain text) |
| Seller name | Not provided for | Required (seller_name) |
Both systems want the same attributes, just in two different containers. For a brand manager, the brand row is worth a closer look: at Google, brand is recommended, in the OpenAI feed it is required.
However, the same OpenAI specification states that the standard upload currently targets the US market; other countries can only be used once OpenAI confirms the integration. For the German market, markup and Merchant Center therefore remain the more important channel for now. Google explicitly recommends maintaining both in parallel, because this helps Google understand and verify the data better (Google Search Central, Product structured data).
My thesis: for D2C manufacturers, GEO is above all product data management, and only after that does content turn this data into answers. At the knife manufacturer, the discrepancies arise where they arise at many manufacturers: the shop admin maintains the product page, the agency builds the Merchant Center feed from last year’s export, and the retailers get an Excel list. And the press release contains the values of the previous model. Four variants of the same data, and the model picks one of them.
That is exactly the argument for a PIM system as the master source: maintain the attributes cleanly once, then push them out to JSON-LD, Merchant Center, retailer portals and, if OpenAI opens the feed for Germany, to the OpenAI feed as well.
Measuring your brand’s AI visibility
Google Search Console now shows you which of your pages collect impressions in AI Overviews, AI Mode and Discover, but neither the clicks nor the query behind them. Google launched the report in June 2026 and has rolled it out to all websites since 31 August 2026 (Google Search Central Blog, 3 June 2026).
Google Analytics 4 does not separate out clicks from an AI Overview either: the default channel “AI Assistant” counts ChatGPT, Gemini, Copilot and the like, while Google’s AI features continue to run under “Organic Search” (Google Analytics Help, Default channel group). So for now, clicks from Google’s AI features cannot be measured separately with the built-in tools.
Neither tool shows whether AI names your brand when someone asks about your category. For that, you still need the spot check from the article on measuring AI visibility, with one extra column for manufacturers: who is named or cited, the brand, a retailer or a marketplace?
For the knife manufacturer, one row would look like this: “Good chef’s knife up to €150 / ChatGPT / brand named: no / cited: Retailer B, review portal”. That sounds homespun, and it is. Even so, it is the only measurement that directly answers the question “Are we being recommended?”
First steps for one quarter
The order I recommend does not start with content, and measurement runs from week one so that you have a before. Played through with the knife manufacturer:
Weeks 1 to 3, reachability. The first level: robots.txt blocks OAI-SearchBot because a blanket bot block from an earlier security update is still active, and the product pages load the technical data via JavaScript. Both are fixed, and Merchant Center gets the current export. If the shop runs on Shopware, the technical levers for this are in the Shopware GEO guide.
Weeks 4 to 8, data foundation. The product markup is validated against Google’s required and recommended fields: 40 of the 400 products are missing a GTIN, and for 120 the hardness rating differs from the PDF. After that, markup and Merchant Center feed are generated from the same source, the markup groups variants via isVariantOf, and the retailers receive the cleaned-up dataset.
The manufacturer sets up the OpenAI feed from the same source at the same time (with variants grouped via group_id there), so that it is ready if OpenAI opens up further markets. If you hand the topic over to a GEO agency, ask them in the initial consultation to show you which source the markup and feeds will be built from.
Weeks 9 to 12, content. For each product category, a guide is written that answers the questions retailers and review portals answer today, such as “Hardness ratings explained”, with the same table and the same figures as on the product page. In week 12, the knife manufacturer asks the same question as at the start. Whether its brand is named then is up to the model. If it is named, the model finds the same hardness rating in the shop, the feed and the retailer data.
Who owns the AI shelf
I think the panic about falling click-through rates is only half justified. Seer’s data shows that the organic click-through rate beneath AI Overviews has been rising again since the low in December 2025, and Google’s own position is that no special optimisation is needed. A properly run website already meets most of what Google requires technically.
The concern is justified at a different point. As a manufacturer, you were always one option among retailers in the search results. In the AI answer, you are either one of the cited sources or you don’t appear at all. That position is won by the brand whose data reads the same in its own shop, at the retailer and on the marketplace. It is the product data management that many manufacturers have been putting off for years, now with a price tag attached.