No more first place, only the source inside the answer
The technical basis has not changed. What is crawlable, cleanly structured and delivered quickly stands the better chance with answer engines too. If you are invisible on Google, ChatGPT will rarely find you. What is new is the target: there is no position 1 left to fight over. There is only the question of whether your paragraph appears as the source in a generated answer — or your competitor's.
That changes what success means. Clicks go down when the answer is already on the results page. The clicks that remain come from people who know what they want; the model has done the shortlisting. That is awkward for reports that count nothing but sessions. It is good for enquiries. This service adds to our SEO strategy, it does not replace it: no technically sound website, no citation.
Answer first, reasoning after
A model rarely quotes a whole page. It takes one paragraph. So every paragraph has to stand on its own: name the question, answer it in the first sentence, explain afterwards. Hide the point in sentence five and nobody reads that far. We write sections that still hold when they are torn out of context — with the company name, the location and the technical term spelled out, because context is exactly what a quotation loses.
Then there is precision. Concrete figures, names, places and conditions can be quoted; general statements cannot. “We offer comprehensive solutions” is the same thing as an empty paragraph to a language model — there is nothing in it to check or repeat. We replace sentences like that with facts someone can look up: delivery time, opening hours, material thickness, system version. The copy gets shorter and less comfortable. And more usable.
Machine-readable: Schema.org, llms.txt, bot access
Structured data is your page translated into machine terms. Product and Offer with price, availability and shipping conditions, FAQPage for direct answers, Organization and LocalBusiness for the company facts. Add a clean heading hierarchy, tables for anything comparable, and an llms.txt as a curated entry point that tells the models which content carries weight. We built this website on the same principles — which makes it our own test case.
Before any of that comes a mundane question: can the bots get in at all? Many shops use robots.txt or a firewall to lock out precisely the crawlers they want to be found by — GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended, Bingbot. That is a decision, not an accident, and it needs to be made bot by bot: if you want to keep training crawlers out but stay in the answer index, you have to tell them apart. We check that first.
Product data, entities and honest measurement
In a shop, the product data is what answers are built from. A model recommends nothing it cannot describe. Complete attributes, dimensions, materials, price, availability, shipping conditions and return rights — properly marked up with Schema.org, not as running text in a “Details” tab. A product photo and three lines of description will not appear in any comparison. Reviews and figures that are comparable across the whole range help as well, because answer engines like to put things side by side. In Shopware 6 we implement this directly in the system.
Models assemble their picture of you from several sources. Company name, location, services and responsibilities therefore have to match beyond your own site: legal notice (Impressum), Google Business Profile, industry directories, trade portals. Contradictory entries usually mean you are not named at all. As for measurement: citations can be sampled, referrals from AI services can be evaluated, brand mentions can be watched. None of them can be guaranteed. Anyone selling you positions in AI answers is selling you a guess.