AI Grounding · Observed from the traffic

What ChatGPT does before it answers "best X".

You type one prompt. Before replying, ChatGPT expands it into several searches, fetches dozens of pages through commercial scraping services, then searches again on what it just read, looping through a few waves before it stops. It keeps most of those pages in a working pool, cites only a handful, and then names a separate set of businesses in the answer. The headline is the gap between how much it reads and how little it cites. Below is one answer, traced end to end.

One answer, end to end

Retrieval pipelines

Every fetched page is tagged (via result_source) with the service that retrieved it. The scraped tiers (bright, oxylabs) are what gets cited; the licensed tier (labrador) is fetched but not cited.

The funnel

Every step

Six stages, top to bottom, for this one answer. Each panel shows the count; expand it for the underlying rows. The last step is not the narrow tip of a funnel: the answer names a separate set of businesses, distinct from the pages it cited.

Stage 1Fan-out: 1 prompt becomes rewritten sub-queries›

Stage 2Search waves: 2 rounds, not one trip›

Stage 3Candidate pool: pages fetched›

Stage 4What the answer uses: most of the pool is never cited›

Stage 5Cited: only distinct pages›

Stage 6Named: the answer names its picks, a separate output›

For those who want to go deeper

The authoritative tier: fetched, retained, not cited

Among the fetched pages is a licensed / authoritative tier the traffic labels labrador: major publishers (AP, Reuters, The Guardian, WSJ, FT, Axios), Wikipedia, and academic sources (arXiv). Across many prompts it is consistently fetched into the pool and consistently not cited. In a live capture, 14 of these sources were present in the model's final consolidated result set and none were cited, so they are retained in the working pool, not filtered out early, then declined at the citation step. The visible answer is sourced almost entirely from the scraped commercial tier.

This answer's licensed pages (fetched into the pool, not cited):

Composition is intent-matched (health queries pull medical and news publishers; B2B queries pull academic and trade sources), but the licensed results are frequently topically off, e.g. a journalism-award story or an unrelated arXiv paper surfacing for an agency query. "Intent-matched" describes the broad domain, not relevance to the specific question.

Hypothesis

ChatGPT may be using the licensed tier as an internal grounding or validation signal rather than a citable source, pulling known-reliable material to anchor its answer and sanity-check the scraped pages it does cite.

A simpler competing explanation

The licensed pages may just be fetched as candidates and then left uncited because they are off-topic for a recommendation answer, with no special grounding role. Because the licensed results are often topically irrelevant, we cannot distinguish these two from the traffic alone.

The honest limit: we can see that the licensed tier is fetched and not cited. We cannot see why; the ranking and selection logic stays on OpenAI's servers. This is a hypothesis to test, not a finding.

Observed, not inferred

The model narrates its own strategy

On the thinking model, ChatGPT emits its own reasoning steps (from the reasoning_title / thoughts stream) while it works. It visibly decides to check vendors for bias and conflicts of interest while researching them by name. This is observed text, not inference, and it is the most concrete evidence here.

Read the numbers carefully

Caveats

The named picks are parsed from prose, not a data field. They need human review; an automated parse has already mis-identified a top pick in our testing. The named set is flagged human-verified.
This is one representative capture; the shape is what repeats. Pool size, cites, and names vary run to run, so the specific numbers here are just this answer. The shape (fan-out, waves, fetch, small cited set, separately named picks) and the licensed-tier-fetched-but-not-cited pattern held across every capture.
The pool is only what reached the browser. Pages the engine fetched and discarded server-side never appear, so any "fetched" count is a floor, not a ceiling.
We cannot see page contents, only each result's snippet. So we cannot claim where the named picks came from (the fetched list pages vs. the model's own training), only that the citation count is far smaller than the named set.
The grounding explanation is a hypothesis with a simpler competing explanation presented alongside it.