Site Search and Recommendations: The Internal Shadow of AI Shortlists

A DTC air-purifier team selling into the US and Germany argues over next month’s content budget: a comparison page, or three explainers on filter fit. The growth lead wants an AI visibility tool to settle it. The technical lead exports ninety days of site-search queries instead. The export opens: “will it help with cat hair,” “how many square feet,” “will the filter fit the model I bought last year,” “does it need the app.” None appear in the tool’s prompt library.

Site search queries, zero-result searches, and co-purchase pairs are the cheapest sample of AI shortlists you already own — a shadow cast only by buyers who arrived, so calibrate it against external prompts before spending.

One job, two retrieval systems

Site search and an AI shopping answer run the same job against different corpora: a sentence goes in, three to five names come out, ranked by match and by why anyone should trust them. Yours draws on your catalog; the assistant draws on the open web and marketplace listings.

That difference sets the limit: the log gives the shape of the question — comparison phrasing, scene constraints, the attribute buyers keep reaching for — not its size in the market.

Co-purchase covers the other half: a model choosing between products needs a swap-in set, and co-purchase is your closest in-house signal for who gets compared with whom. Accessory pairs decide whether a compatibility answer can be built.

The bias is structural: treating your query distribution as the market’s is the most common mistake.

Three data types, one usable

Query text carries the most information, zero-results the most urgency, co-purchase the most structure. In one “top searches” report, all three go stale.

Data	Read it for	Typical failure
Queries	Buyer phrasing: comparison, scene, attribute	Reporting a no-result rate, not sentences
Zero-result, low-click	Questions you cannot answer, or answer badly	Filed under “other” and closed
Co-purchase	Who is compared with whom, what sells together	Recommendation click-through only

All three join on one key: query → page → product or category. Log the raw query, result count, first result, and click. Drop one and “we could not answer” versus “we answered and nobody looked” becomes a guess.

Five steps from query to content gap

The output is a table with an owner and a review date on every row, not a report.

Step	Action	Owner	Passes when
1 Normalize	90 days; dedupe, merge variants, drop brand and navigation terms	Analytics	Unique queries fall, clusters hold
2 Tag	Comparison, scene, attribute, post-purchase	Content, ops	Nine in ten queries land in one type
3 Map	Point each cluster at a URL	Content	Every cluster has a page or is marked “none”
4 Diagnose	No page is a gap; a page with no clicks is a wrong answer	Content, front end	The two lists stay separate
5 Verify	One owner and one recheck date per cluster	Growth	Cluster clicks reviewed at 30 days

Step 4 needs the judgment. A query with results and no clicks usually has a title and snippet that fail the sentence, not a ranking problem. A zero-result query is usually a missing synonym, not a missing page: mapping “purifier filter” to “replacement filter” is an hour against a month. Either way the cluster must compete for a seat, which is how AI shortlists only 3–5 names.

Example: ninety days that changed a budget

Example — brand and counts invented; the method is the point.

Say ninety days produce 18,000 searches and 4,200 unique queries after normalization. Tagged, scene queries sit near half, attribute second, comparison under a fifth. The repeated lines: “how many square feet,” “how long for a new build,” “cat hair,” “are filters cross-compatible.”

The team believed buyers compared specs; the log showed them confirming fit for a specific room. Budget moved from a discount campaign to three assets: a filter-to-model compatibility table, a room-size guide stating its airflow assumption, and a pet-household note. Co-purchase agreed: machine buyers add filters, and unclear compatibility breaks the bundle.

They tracked one number: whether those clusters moved from page two of internal results to page one. That number proves itself; conversion cannot separate the page from the discount beside it.

The return was one corrected assumption, not three topics. Cluster edges also trace the category tree AI can understand.

Calibrate against external prompts

Site data decides what to make; external prompts whether it is worth making. When they disagree, do not force priority out of the log.

Rewrite the top fifty clusters as questions for an assistant — strip brand words, add the scene, ask twice in two or three entry points. Record three things: are you in the shortlist, who else appears, how you are described. Sample rather than screenshot, as an AI visibility baseline does.

Heavy overlap: schedule it first; the shadow is trustworthy and returns most per hour.

Site-heavy, assistant-light: demand has already arrived. That is a conversion problem — product page and review corpus, not content budget.

Site-light, assistant-heavy: the dangerous case. Buyer phrasing has moved ahead of your traffic mix, and the log cannot see it. External prompts lead; the log only validates.

Where this breaks

The sample is too thin. A few hundred searches a month, two or three per cluster, supports no extrapolation. New stores and markets start from external prompts.

The box is navigation. Brand names, SKU codes, “track order.” Fix the results page before reading the log.

Sales sit in marketplaces. Those queries are not yours and the console gives aggregates; the log covers the DTC slice only.

Compliance forbids it. Raw queries can contain emails, phone numbers, delivery addresses. Redact before export; treat cross-border transfer and third-party upload under local rules.

The category is high-consideration. Research happens in communities and video; site search reflects people who already decided. It optimizes conversion, not shortlist prediction.

Instrumentation, join keys and redaction belong to the technical foundation; topic judgment to content. Split them and the export rots on a drive.

This month: export ninety days of site search, redact and normalize it, tag the four types, take fifty clusters. Rewrite ten as questions for an assistant, ask each twice in two entry points, and record only whether you appear. One person, two days.