A shopper asks for "a quiet cordless vacuum for hardwood and pet hair under $300." The assistant does not run one search. It fires a dozen, reads the passages it gets back, grades its own first draft, then goes back for more if the evidence is thin. Your product can get retrieved and still lose, because a second pass decided a rival's data answered the question better.
TL;DR: AI shopping search has moved past single-shot retrieval. Google AI Mode, ChatGPT Search, and Perplexity now plan, retrieve, read, grade their own answer, and retrieve again before they respond. One shopping query triggers five to twenty internal sub-retrievals, and your product passages get compared head-to-head against rivals at each step. Being mentioned once is no longer the bar. Surviving the grading loop is.
This is the part of AI search almost no commerce team is measuring. Most brands still ask one question: "do we show up in ChatGPT?" That checks the final answer. It tells you nothing about the four or five upstream stages where your product was either pulled in or quietly dropped. As of June 2026, the trade press has finally named the architecture that runs underneath every major answer engine, and it changes what "AI-ready product data" has to mean.
What Changed: From Retrieve-Once to Retrieve-Read-Retrieve
The old pipeline was a straight line. A query came in, a vector index returned the top matching passages, the model read them, and it wrote an answer with citations. If your page was in that first pull, you had a shot. If not, you were invisible. That linear model is now obsolete across the major platforms.
Search Engine Land, reviewing how the architecture evolved since the 2020 retrieval-augmented generation research, describes the shift bluntly: "The retrieve-once-then-generate pattern that defined the first wave is obsolete." The new pipelines add four properties the straight line lacked: planning, tool routing, multi-hop iteration, and reflection. A single user query now triggers "somewhere between five and twenty internal sub-retrievals," and the agent only writes a final answer "once it has decided the evidence base is sufficient."
For a commerce team, the practical meaning is simple. The assistant does not just look for your product. It looks, judges what it found, and decides whether to look again. That judgment step is where most catalogs lose.
The Grading Step Nobody Audits
The single most important new behavior is reflection: the agent grades its own draft answer and decides whether the evidence is good enough. This is the gate that rejects product data quietly. You never see the rejection. You only see whether you ended up in the final response.
Google described the front of this loop at I/O 2025 when it launched AI Mode: the system "breaks down your question into subtopics" and "issues a multitude of queries simultaneously." That is the well-documented query fan-out stage, which we covered in detail in our hidden retrieval layer breakdown. What the newer architecture adds on top is the reflection pass: after the fan-out runs, the agent reads the passages, compares them, grades the synthesis, and re-retrieves when a sub-query came back weak.
In agentic retrieval you cannot see the gatekeepers rejecting you. You only see whether you ended up in the final answer.
That is the whole problem in one sentence. 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, is a black box you can only probe by testing the surfaces that consume your data. This is why passage-level retrieval matters more than page-level ranking: the unit being graded is a passage, not a URL.
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. A query like "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 title, price, and a marketing paragraph, you answer one sub-query and miss four. The agent stitches together an answer from competitors who covered the rest.
A weak first pull triggers a re-retrieval that finds your rival. When the initial pass misses the canonical spec, the agent does not give up. It searches again, often with a sharper sub-query, and that second search surfaces whichever brand published structured, machine-readable attributes for that exact dimension.
Modality mismatch loses the comparison. Comparison questions favor passages structured as tables and spec lists. Process questions favor numbered steps. A product described in prose competes badly against a rival whose attributes are laid out in a format the agent can lift directly into its answer.
| Stage | Old single-shot RAG | Agentic retrieval (2026) | What it rewards |
|---|---|---|---|
| Query handling | One query, one retrieval | One query splits into 5-20 sub-retrievals | Coverage across sub-queries |
| Evidence pull | Top-k passages, once | Retrieve, read, re-retrieve on weak pulls | Depth per attribute |
| Selection | Whatever was in top-k | Head-to-head passage grading | Structured, extractable data |
| Output | Answer from first pull | Answer only after self-grading | Surviving the reflection pass |
The Stat That Explains the Invisibility
Most brands assume that ranking on page one of Google means AI will cite them. The data says otherwise, and it is the clearest evidence that the grading loop runs independently of classic rank.
68% of pages cited in AI Overviews do not rank in the top 10 organic results for the main query. Pages that rank for the main query and its fan-out sub-queries are 161% more likely to be cited.
That finding comes from a Surfer SEO study of 173,902 URLs. The pages getting pulled into AI answers are the ones that answer the specific sub-queries the agent generates, not the ones that won the head term. Surfer measured a Spearman correlation of 0.77 between the number of fan-out sub-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-queries that actually decide the answer ("stability running shoes for overpronation under $140 with a wide toe box").
This is the commerce edge of retrieval-augmented generation. The brands winning are not the ones with the best homepage. They are the ones whose product attributes survive grading across the whole fan-out.
What This Means for Mid-Market Brands
Enterprise retailers have data teams who can chase this. Mid-market brands usually do not, which is exactly why the gap is widening every month. The good news: the fix is structural, not a budget arms race. It is about how your product data is shaped, not how much you spend.
The brands that win the second pass treat each product attribute as something an agent must be able to retrieve, read, and lift into a comparison without guessing. That means entity-level clarity about what your product is, who it is for, and how it compares, expressed in formats the agent can extract. It also means measuring your product found rate across sub-queries, not just the top-level "do we appear" check.
What to Do This Week
- Run a fan-out audit on your top category. Take your three highest-revenue category queries, expand each into the eight to twelve sub-queries a shopper would actually need answered (price band, use case, fit, materials, compatibility), and check whether your product data answers each one. Tools like our AI Readiness Report score this gap in about 30 seconds.
- Convert prose to structured attributes. Pick your 20 top products and move the buried specs out of marketing paragraphs into discrete, labeled attributes. The agent reads attributes, not adjectives.
- Cover the comparison dimensions. For each product, list the three attributes a shopper compares against rivals (noise, runtime, weight, fit). Make sure each is present and machine-readable, not implied.
- Test the surfaces directly. Since the upstream stages are a black box, the only way to see the grade is to query ChatGPT, Google AI Mode, and Perplexity with real shopper questions and watch whether you survive to the final answer.
- Track it over time, not once. The grading loop shifts as models update. A single snapshot is a guess. Weekly tracking against the same sub-queries shows whether your fixes are landing.
Frequently Asked Questions
What is agentic retrieval in AI search?
Agentic retrieval is the multi-stage process AI search engines now use instead of single-shot retrieval. The system plans an answer, breaks the query into sub-queries, retrieves passages, grades its own draft, and retrieves again if the evidence is weak. A single query can trigger five to twenty internal sub-retrievals before any answer is written.
How is this different from query fan-out?
Query fan-out is one stage: the decomposition of a query into parallel sub-queries. Agentic retrieval wraps fan-out inside a larger loop that also includes reflection, where the agent grades its draft and re-retrieves on weak results. Fan-out determines what gets searched; the grading loop determines what survives.
Why does my product show up sometimes but not always?
Because the agent grades the evidence each time and the sub-queries vary by phrasing. If your data answers some sub-queries strongly and others weakly, you survive the grade for some shopper questions and get dropped for others. Consistent visibility requires coverage across the full fan-out.
Does ranking on Google page one guarantee AI citation?
No. A Surfer SEO study of 173,902 URLs found 68% of pages cited in AI Overviews do not rank in the top 10 for the main query. AI engines pull passages that answer specific sub-queries, which is a different signal from classic head-term ranking.
What kind of product data survives the grading pass?
Structured, discrete, machine-readable attributes that map to the dimensions shoppers compare: price band, use case, materials, fit, compatibility, and performance specs. Prose marketing copy answers the head query but loses the passage-level comparison against rivals who expose the same facts as extractable data.
The shift from retrieve-once to retrieve-read-retrieve is the most consequential change in shopping search since mobile-first indexing, and it is already live across every major answer engine. The brands that treat AI visibility as a single "do we show up" check are watching the last frame of a movie they never saw. The ones measuring whether their product data survives every grade in the loop are the ones AI keeps recommending.
How AI-ready are your products?
Check how ChatGPT, Google AI, and Perplexity evaluate any product page. Free score in 30 seconds.
