The prompt as a Lego figure: why AI visibility runs on entities, not keywords

Most optimization plans I see in consulting projects still follow the same logic: one prompt, one landing page, one ranking. That equation no longer works, because ChatGPT, Google AI Mode and Perplexity break every question apart before answering. A prompt is then no longer a single question. It is a Lego figure made of many blocks, and the AI searches for each block on its own. A good Google position on the starting term can become a bit part in the AI answer — or vanish entirely.

How SEO shifted here in four stages

I still remember the stage of pure keyword stuffing. A block of comma-separated terms at the bottom of the page was often enough for stable positions, without our having to move much else. After that came research with intent separation: informational, commercial, transactional. Semantic cores, clusters, topic maps. The transition to the era of language models — heralded by Google’s BERT, which reads sentences in context rather than word by word — rewired the model once more. Search engines understand context, prepositions, irony. The exact term loses its rank, semantic depth takes its place.

Now comes the fourth stage. The question of which keyword a page should rank for is replaced by another: which sub-questions the AI search process derives from the starting term itself, and in which of them your content appears. That is not a gradual refinement. It is a new foundation.

What Google officially calls query fan-out

Elizabeth Reid, Head of Search at Google, named the mechanism at Google I/O 2025: “Search recognizes when a question needs advanced reasoning. It calls on our custom version of Gemini to break the question into different subtopics, and it issues a multitude of queries simultaneously on your behalf.” So the system recognizes when a question goes beyond a simple answer, breaks it into subtopics and fires a separate search for each part. The final answer builds on this bundle, not on a single ranking.

How large this bundle gets depends on the reasoning mode. Kevin Indig and Semrush, in an analysis from June 2026, ran 100 prompts through GPT-5.2 in both modes and compared the sub-queries. Thinking Mode fires 4.6 times as many fan-out queries as Instant Mode. Only 25.6 percent of the cited domains overlap between the two modes. Almost three of four pieces of evidence never appear in the fast mode. So whoever measures their visibility only in Instant Mode measures past the competitive field a buyer actually sees on a complex question.

A case from my practice: “Is fulfillment outsourcing worth it for my shop?”

An example from an ongoing fan-out analysis for a Swiss e-commerce context. The prompt “Is fulfillment outsourcing worth it for my shop?” broke into nine sub-queries the system fires in parallel. Seven of them revolve around “fulfillment outsourcing benefits for online shop.” The system wants to solve the core question from the angle of pros and cons. Two additional sub-queries break the core question down further. The first opens the buying environment: “fulfillment outsourcing e-commerce worth it Switzerland cost benefits drawbacks.” The second pulls in platform and provider landscape: “fulfillment outsourcing benefits drawbacks e-commerce worth it fulfillment provider Germany.”

If I extract the most frequent word pairs from this and further repetitions of the fan-out, it becomes visible what the buyer is pointing at in the decision process. The distribution looked like this: “benefits online” leading with 28 mentions, then “e commerce” with 18 and “benefits drawbacks” with 17, followed at a distance by “commerce benefits” with 6. The long tail after that consists of single hits on Swiss costs, Amazon FBA, warehouse logistics and provider country, one or two mentions each. The prompt reads like a single question. It is in reality a bundle of checks in which the buyer probes geographic, price, technical and competitive boundaries.

For a fulfillment provider from Switzerland these are measurable visibility positions. If a clear, citable answer on “costs in Switzerland” or on the comparison with Amazon FBA is missing, the model gets the answer elsewhere. The “fulfillment outsourcing” landing page may sit at position 3 on Google. For the ChatGPT buyer who calculates costs in francs and has FBA in mind as an alternative, it is invisible. The case is still running. Which gap ends up bringing how much citation rate I will add once the new answers are indexed — result to follow.

What 25 prompts reveal about the fan-out

A single prompt can mislead. So I laid the same lens on the whole set: 25 buying prompts from this Swiss fulfillment context, tracked across ChatGPT and Perplexity, together 1,390 fan-out sub-queries in a one-month window. Three findings from it change how I plan visibility.

The first number is the breadth. No prompt stays in its starting form. At the median, every question breaks into 26 sub-queries, the narrowest into 19, the broadest into 31. The fan-out thus reaches further than the nine from the single example above suggest — across two engines combined, it is the normal case.

The second number carries the thesis. For 306 domains I laid two values side by side: in how many of the 25 prompts a domain is cited, its coverage breadth, and how often the engines retrieve it at all, its retrieval share. The two are strongly related, Pearson correlation +0.64. Whoever appears across many sub-queries is drawn on noticeably more often. The counter-check sharpens the point: against citation density per hit, breadth does not correlate at all (−0.05). What decides is not how densely you are cited in one spot, but on how many blocks of the figure you appear.

The third finding shows what the fan-out asks about. Every sub-query probes a boundary of the buying decision. Most often geography: 58 percent name a country or economic area, Switzerland, Germany, UK, EU. In 41 percent there is a comparison — “best,” “top,” direct provider check. Just under a third checks hard criteria and regulation, from customs clearance to scalability. Cost appears in 8 percent, explicit alternatives in 2. The buyer asks one question and means a dozen checks.

Why brief-per-prompt thinking fails here

The classic workflow looks like this: we want to be visible for a certain prompt, so we write a brief that tailors a page to that prompt. The workflow falls short, because the prompt is not the optimization object at all. It is a Lego figure, assembled from many individual blocks. Each of these blocks is a fan-out query. Whoever builds a good page for the figure but ignores the individual blocks has only hit the silhouette. The blocks it is assembled from come from other sources.

The practical consequence is uncomfortable: a good SEO position for the starting term and a visibility in the AI answer are two separate quantities. One value can rise while the other falls. Checking both is mandatory.

Entity mapping as the new foundation

From this shift follows a different foundation. Instead of building one page per prompt, you map the entities of your market and assign your answers to them. An entity, in the context of Generative Engine Optimization (GEO), is a clearly bounded object: “fulfillment provider Switzerland,” “Amazon FBA,” “Shopify fulfillment,” “cost per parcel in CHF,” “return rate.” Each of them can land in a fan-out query. If your pages evidence these entities with clear, verifiable statements and link them cleanly to one another internally, your domain becomes the obvious source for the model.

The value of this way of thinking carries further than the single prompt. A system that answers in fast mode today can insert the longer reasoning step tomorrow. The models will research more deeply along the way, but the entities and their relationships remain. Whoever deposits them clearly is cited in both modes. This matches the same principle Google confirmed at Search Central APAC 2025: PageRank lives on, and search engines think in entities and their relationships, not in isolated keywords.

Fan-out check: the workflow for your next three prompts

A recipe that works without a new tool.

First: pick three prompts that, in your view, map your most important buying decisions. A good test: would you buy on exactly this prompt yourself?

Second: have the model show the fan-out. Phrase it: “Show me the sub-queries you would internally run for this prompt, with a short justification.” ChatGPT in Thinking Mode and Perplexity answer this reliably. Note between ten and fifteen sub-queries per prompt.

Third: extract the most frequent word pairs. Expressions like “price per parcel,” “provider Switzerland” or “comparison fba” show you the boundaries at which the buyer checks. Hold them in a lean table, word pairs in rows, prompts in columns.

Fourth: for each word pair, check whether your content contains a clear, citable sentence on it. One sentence, no essay around it. The sentence has to stand so the model can lift it out and use it as evidence. If the sentence is missing, the gap is found.

Fifth: sort the gaps by two measures — importance for the buying decision and effort for the answer. Fill the gap with the best ratio first.

The whole workflow for three prompts costs me, in practice, between half a day and a day. Measurable movement in the citation rate — the share of answers in which the model names you as a source — usually shows within a few weeks, once the new answers are indexed.

Where this approach reaches its limits

Not every prompt is worth a fan-out analysis. For prompts that depend heavily on real-time prices or day-to-day news, the bundle of sub-queries shifts every week. The effort of maintaining every shift exceeds the visibility gain. The fan-out check pays off for prompts that map a structural buying decision: several criteria, longer research, a clear provider comparison.

A second point: entity mapping does not replace clean technical SEO. If GPTBot, ClaudeBot or PerplexityBot cannot reach your pages, every mapping stays worthless. How status codes and crawl budget provide this foundation I described in the piece on crawl budget.

Conclusion

The old SEO equation prompt → landing page → ranking carried for a long time. It carries no more, because between prompt and answer sits a new step that neither users nor classic ranking tools see: the fan-out into sub-queries. Whoever reads the sub-queries sees the buyer’s boundaries clearly. Whoever evidences the entities behind them with clear answers is cited in both reasoning modes. Whoever keeps writing for the starting term optimizes for a stage that no longer exists in that form.