Why ChatGPT Recommends You on One Question and Skips You on the Next

A shopper asks for "a quiet cordless vacuum for hardwood and pet hair under $300." The assistant does not run one search. It runs several, reads the passages that come back, grades its own first draft, and goes back for more when the evidence is thin. Your product can be retrieved and still lose, because a second pass decided a rival's data answered the question better.

That is the mechanism behind a question we hear from ecommerce leaders constantly: why does ChatGPT recommend us on one question and skip us on the next? Most teams check one thing, whether they appear in the final answer. That tells you nothing about the upstream stages where the product was pulled in or quietly dropped.

What changed: retrieve, read, then retrieve again

The first wave of AI search was a straight line. A query came in, an index returned the best matching passages, the model read them and wrote an answer with citations. If your page was in that first pull, you had a shot. If not, you were invisible.

Michael King, writing in Search Engine Land in late May 2026, describes how every major answer engine has moved past that model. The new systems plan, route between tools, retrieve, read, then retrieve again, and grade their own drafts before deciding whether to look further. In his words, "the retrieve-once-then-generate pattern that defined the first wave is obsolete."

The same analysis puts the scale of it plainly: a single user query triggers "somewhere between five and twenty internal sub-retrievals," and the agent only writes a final answer once it decides the evidence is sufficient. For a commerce team the meaning is simple. The assistant does not only look for your product. It looks, judges what it found, and decides whether to look again.

The grading step you cannot see

The most important new behavior is reflection. The agent grades its own draft and decides whether the evidence is good enough. This is the gate that rejects product data quietly. As King puts it in the same piece, "you cannot see the gatekeepers rejecting you. You only see whether you ended up in the final answer."

Rank checking, citation counting and prompt by prompt sampling all observe the last stage of a multi stage process. Everything upstream, the planning, the sub-retrievals, the grading, can only be probed by testing the surfaces that consume your data with real shopper questions and watching what survives.

The front of this loop is well documented. Google described AI Mode at I/O 2025 as a system that breaks a question into subtopics and issues "a multitude of queries simultaneously". That is query fan-out, which we explain in query fan-out for commerce. Fan-out decides what gets searched. The grading pass decides what survives, and it runs again whenever a sub-question comes back weak.

Why product data loses the second pass

Three failure modes show up over and over when product data survives the first retrieval and dies on the grade.

Compound questions split your data

"Quiet cordless vacuum for hardwood and pet hair under $300" needs separate retrievals for noise level, floor type, pet hair performance, cordless runtime and price band. If your product page carries only a title, a price and a marketing paragraph, you answer one of those and miss the rest. The agent stitches together an answer from competitors who covered them.

A weak first pull triggers a re-retrieval that finds your rival

When the initial pass misses the spec it needs, the agent does not give up. It searches again, often with a sharper sub-question, and that second search surfaces whichever brand published a clear, machine readable answer for that exact dimension.

Prose loses the comparison

Comparison questions favor passages laid out as spec lists. Process questions favor numbered steps. A product described only in prose competes badly against a rival whose attributes sit in a format the agent can lift directly into its answer.

Put side by side, the two systems reward different things:

• Single shot retrieval ran one query once. Agentic retrieval splits it into many sub-retrievals, so coverage across sub-questions wins.
• Single shot retrieval took the top passages and stopped. Agentic retrieval re-retrieves on weak pulls, so depth per attribute wins.
• Single shot retrieval used whatever came back. Agentic retrieval grades passages head to head, so clear, extractable data wins.
• Single shot retrieval answered from the first pull. Agentic retrieval answers only after self grading, so the products in the answer are the ones that survived the second look.

Why page one of Google does not carry over

Most brands assume that ranking on page one means AI will cite them. In Surfer SEO's study of 173,902 URLs, roughly 68% of pages cited in AI Overviews did not rank in the top 10 organic results for the main query or for any of its fan-out queries. Pages that ranked for both the main query and its fan-out queries were 161% more likely to be cited than pages ranking for the main query alone.

Surfer also measured a Spearman correlation of 0.77 between the number of fan-out queries a page ranks for and its odds of being cited. For commerce the read is direct. A catalog tuned for one head term, "running shoes," is invisible to the sub-questions that decide the answer, such as "stability running shoes for overpronation under $140 with a wide toe box."

What this means for a lean ecommerce team

The fix is structural, not a budget race. It is about how your product information is shaped, not how much you spend. Each attribute a shopper compares on has to be something the agent can retrieve, read and lift into a comparison without guessing, and it has to be present for every product that matters, not only the hero SKU.

When a Paz expert looks at a catalog with this problem, the work comes down to three questions:

1. Which shopper questions do you survive? Take the questions real customers ask in your category, phrased the way they would say them to an assistant, and run them in ChatGPT, Google AI Mode and Perplexity. Record whether you appear as a product card, a mention or not at all.
2. Which attributes are missing where you lose? For each question you fail, list the sub-questions behind it (price band, use case, fit, materials, compatibility, performance) and check whether the page states the answer in plain text or buries it in an image, a PDF or a support article.
3. What does before and after look like? Fix the gaps from your own product sources, then re-run the same questions on a schedule. The grading loop shifts as models update, so a single snapshot is a guess. Weekly reads against a fixed question set show whether the fixes landed.

If you want a quick first read on a single product page, the free AI readiness report shows what the assistants can and cannot find on it.

Frequently asked questions

What is agentic retrieval in AI search?

Agentic retrieval is the multi stage process AI search engines use instead of single shot retrieval. The system plans an answer, breaks the query into sub-questions, retrieves passages, grades its own draft and retrieves again if the evidence is weak. Search Engine Land puts the range at five to twenty internal sub-retrievals for one query.

Why does my product show up sometimes but not always?

Because the agent grades the evidence each time and the sub-questions change with phrasing. If your data answers some sub-questions strongly and others weakly, you survive the grade for some shopper questions and get dropped for others. Consistent visibility requires coverage across the whole set.

Does ranking on Google page one guarantee an AI citation?

No. Surfer SEO found that most pages cited in AI Overviews did not rank in the top 10 for the main query. AI engines pull passages that answer specific sub-questions, which is a different signal from head term ranking.

What kind of product data survives the grading pass?

Clear, discrete attributes that map to the dimensions shoppers compare: price band, use case, materials, fit, compatibility and performance specs. Marketing prose answers the head query and loses the passage level comparison against rivals who expose the same facts as extractable data.

Where Paz fits

This is the work Paz does for lean ecommerce teams. A Paz expert writes the shopper questions in your category, finds the attributes your pages and feed are missing where you get dropped, and fixes the product content from your own sources. You approve each change, and the platform publishes it and tracks the before and after, question by question. If you want to know which questions you are losing on the second pass, talk to us.