On August 8, 2026, ChatGPT went from showing 11.6 pages consulted per answer to 22.5, and from 0.2% to 18% of its internal sub-queries using a site: operator against a named domain. It happened overnight. Regulators and standards bodies gained reading share; affiliate review media lost on both reading and citations. 85.9% of tracked brands are now cited from their own site more often than two weeks earlier. One open question remains: whether ChatGPT does more research now, or merely discloses more of what it always did.
Across 5,000+ tracked brands, the pages ChatGPT shows it consulted went from 11.6 to 22.5 per answer, and it began aiming site: searches at named domains. It happened overnight. Regulators gained most of the reading, consumer affiliate media lost on every measure, and 85.9% of brands are now cited from their own site more often than two weeks ago.
Domain-targeted searches inside the fan-out
Share of ChatGPT's internal sub-queries that use a site: operator to restrict results to one named domain. Flat for months, then a step change that has held for ten days.
Unbranded prompts, counting only sub-queries the model generated itself.
Five numbers that define the August 8 change
| Measure | Before | After | Change |
|---|---|---|---|
| Pages shown as consulted, per answer (paired, same prompts) | 11.58 | 22.54 | 1.95× |
| Answers running multi-query research | 2.0% | 57.2% | 28× |
site: share of fan-out sub-queries | 0.2% | 18% | ~90× |
| Inline citations per answer | 0.49 | 0.82 | 1.68× |
| Brand mention rate (share of voice) | 40.5% | 37.6% | −2.9 points |
Did ChatGPT do more research, or just disclose more of it?
Why this question matters. Fan-out and source lists come from the product's research disclosure panel. If OpenAI simply started showing more of a process it had always run, the source-side results would reproduce with zero change in behavior.
What we can rule out, and what we can't. Two scrapers agreeing rules out one vendor's bug, but both read the same surface. The non-browsing API route is only ~2% of our ChatGPT volume, with the rest coming from provider scraping, so it amounts to a footnote, and it never spoke to the disclosure question anyway.
Why answer length can't settle it. On 287,157 identical prompts, ChatGPT answers did grow, but the median prompt gained only +5.1%, with 56.8% coming out longer. That's barely above a coin flip.
Every other engine moved the same way. Same paired basis, same window: AI Overviews +4.7% median, Gemini +6.9%, Copilot flat at −1.5%, Perplexity −34.5%. Three engines moving together points at something shared across all of them: prompt population, our own extraction, or a shift on the vendor side.
Set the disclosure question aside for a moment, because the rest of this holds whichever way it resolves.
Depth changed. Length within each depth didn't. A three-sub-query answer looks much the same in July as in August. There are simply 40× more of them. The small rise in the mean comes from that shift in composition.
| Fan-out depth | Share of answers before | Share after | Median length before | Median length after |
|---|---|---|---|---|
| 1 sub-query | 98.0% | 42.8% | 2,355 | 2,340 |
| 2–3 | 1.9% | 44.3% | 2,472 | 2,538 |
| 4–7 | 0.04% | 9.9% | 3,111 | 3,306 |
| 8+ | <0.01% | 3.0% | n=1 | 4,580 |
The trail changed shape as well as size. 43% of answers still show a single sub-query, exactly as before, while the rest moved into multi-step. Whether that reflects new routing or newly visible routing is the open question above. Either way, it's what we and everyone we track now observe.
It doesn't hit every prompt equally. Sorted by intent, the effect concentrates on comparison and consideration queries:
| Prompt intent | Sources before | Sources after | Ratio |
|---|---|---|---|
| Commercial ("best X", "X vs Y") | 11.60 | 24.64 | 2.12× |
| Informational | 11.73 | 20.51 | 1.75× |
| Transactional | 11.01 | 17.58 | 1.60× |
Why 478 brands barely moved. They run 39.9% branded prompts against 12.2% for the rest, and skew informational. This isn't a collection artifact: same scraper, same route, same baseline depth. They simply own prompt sets that the change barely touches.
What would settle this, and why we can't run it yet
The test that would settle it. A structural diff of raw captures either side of 00:30 UTC on Aug 8: did the source and sub-query list move to a different element, or did an existing element start filling up?
Why we can't run it. Raw HTML captures aren't retained in the analysis dataset, and the search-query records carry no provenance field. Answering it means going back to the raw captures while they're still inside their retention window.
What holds either way. Every citation-side result stands, because inline citations come from the answer body. So does the timing of the break, which lands inside a running collection window rather than at its edge, and so cannot be our own run changing.
What depends on the answer. Pages read per answer, the site: share, and the read-side category mix. If the pre-period only ever disclosed one sub-query, those denominators are weak. That's why the headline reads "0.2% → 18%" and not a single dramatic multiple.
ChatGPT reads more institutions, but cites more vendors
Every figure below is a share ratio that has already been normalized against pool growth, so 1.00× means a category held exactly the share it had before. Above 1.00× it gained ground, below it lost ground, and that holds even where the raw count rose. The pool grew 1.82× for reading and 1.68× for citing, and the normalization has already removed both, so compare these numbers against 1.00× and never divide by pool growth again.
Read share versus cited share, by source category
Share ratios, 1.00× = held its share. The two bars diverge sharply for brand-owned pages and for social, because a site can gain in one and lose in the other.
Read share
Cited share
The commercially important divergence: brand-owned pages barely gained reading share (0.95×) but gained citation share by 80% (1.80×). Editorial, review platforms and forums lost in both, and lost roughly twice as hard in citations as in reading.
Share change by individual site
Share ratios, already normalized, so 1× is neutral. Log scale, meaning each step doubles or halves. Brand counts are shown so you can see no single brand's data sits behind a bar.
Every site here is computed on one method: fixed brand cohort, same two windows, share-normalized. Sites with larger declines exist (healthline.com at 0.22×, ~28,000 reads across 501 brands) but fall below the volume floor and are excluded rather than shown unqualified.
Where ChatGPT aims its domain-targeted searches
The named targets are review and reputation platforms. ChatGPT interrogates them by name: reddit.com (1,657 brands), trustpilot.com (1,166), gov.uk (650), bbb.org (340), g2.com (295), clutch.co (219).
But the biggest single target is the brand's own domain. 12.2% of all site: queries point there, 113,122 of 927,875 over nine days, rising to 14.6% on a looser subdomain match.
Both things are true at once. Platforms rank individually because they're shared across thousands of brands; own-domain targets fragment across thousands of distinct domains and never chart.
Does ChatGPT know the brands already, or find them first?
Why the order of the queries matters. If ChatGPT issues site:yourbrand.com queries, maybe the brand set is decided in advance from what the model already believes and the searching is just confirmation. If so, retrievability matters far less than being baked into the weights.
What the sequence actually shows. Across 175,098 answers containing a site: query, the dominant shape is one broad query first, then named-domain drilling.
of site:-using answers open with it
No prior search at all. The other 91% run a broad query first, then drill into named domains.
The usual path is discover, then verify. See who's in the field, then interrogate those domains by name. Good news for retrievability: the brands being checked largely surfaced in a search first.
The 8.9% that skips discovery. ~15,500 answers whose first action is querying a named brand's domain with zero prior retrieval. Those names can't have come from results, because there were none. A "knows first" mode exists. It just isn't the main one.
What the query sequence cannot tell us
For the other 91%, we can't tell where the drilled domain came from: the first query's results, or the model's priors. There's no sub-query-to-source link in the data, so the two are indistinguishable. That's the question that decides AEO strategy, and we can't close it with what we collect today. Whether being remembered or being retrievable determines who gets checked stays open, so we claim no more than the sequence shows.
More brands get cited, and the leader takes a bigger share
Two things happened at once. They sound contradictory until you look at the denominator. Across 930 brands with enough citations to rank in both windows:
| Per brand, competitor pages cited inline | Before | After | Change |
|---|---|---|---|
| Distinct brands cited | 28.66 | 35.05 | +22% |
| Total citations | 73.6 | 137.0 | +86% |
| Top brand's share of them | 19.3% | 21.7% | +12% |
| Concentration (HHI) | 0.102 | 0.117 | +15% |
The pie nearly doubled and about six more brands got a slice of each tracked brand's answers, but the leader's slice grew faster than the field widened. The door opened wider, and whoever was already in front gained more.
Citation rank churn is real, and far smaller than it first looks
Our first cut looked spectacular. Only 2 of the top 5 cited brands survived, and 72% of brands had a new number one.
Almost all of it is an artifact of small lists. These are low-count lists with many near-ties, so the top five churns on its own. Comparing two windows that both sit before the change shows the baseline.
Leaderboard churn against a null baseline
Higher bars mean more stability. The null compares two pre-change windows (Aug 1–3 vs Aug 5–7); the test compares across the change (Aug 5–7 vs Aug 9–11). Equal window lengths, same 347 brands throughout.
Null (both windows pre-change)
Test (across Aug 8)
Even with nothing changing, 57% of the top five turns over between adjacent windows and the leader changes 62% of the time. The change adds roughly 12 points of churn: directionally real, modest, and invisible without the baseline.
Own-domain citations rose for 85.9% of brands
The aggregate figure, brand-owned pages gaining 80% citation share, could have been a handful of brands dragging a mean. It isn't. Across 1,665 brands with 100 or more answers in both windows:
of brands increased their own-domain citation rate
Median per-brand gain of 3.48×, or roughly 2.1× after normalizing for pool growth. Strongest single result in the study, and the easiest to act on.
Reddit fell off a cliff in ChatGPT answers on August 14
A separate discontinuity, six days later. Reddit went from 12.6% of ChatGPT answers to 1.7% in ~48 hours and has stayed at the floor. Total sources per answer held at 18.5, so it isn't a general contraction.
Reddit alone, out of every comparable source. Every other UGC and review platform is flat or up over the same days. LinkedIn +54%, Facebook +125%, YouTube +58%, with Trustpilot, G2, Wikipedia and TripAdvisor unchanged. A "downweight UGC" policy would have hit all of them.
ChatGPT kept asking; Reddit stopped answering. Requests mentioning Reddit fell 25%; Reddit pages actually read fell 87%. Yield per request fell to about a sixth of what it was. That asymmetry is an access cutoff, not a ranking change.
ChatGPT kept asking for Reddit. Reddit stopped arriving.
Both series indexed to August 8 = 100 so they share one scale.
Reddit requests in fan-out
Reddit pages actually read
Requests −25%, yield −87%. Reddit is still the domain ChatGPT names most often in its own searches, and it now returns almost nothing.
One caveat to keep attached to the Reddit finding
Magnitude varies by collection provider. On Aug 15 one saw Reddit in 2.18% of answers and another in 5.49%, having agreed within a few points before Aug 14. arXiv shows a similar shape on the same days.
Direction and timing agree everywhere. Treat 12.6% → 1.7% as our best estimate and expect the exact figure to move, and don't build a claim about any named commercial arrangement on top of it.
What the August 8 change means for brands
Fix your own site's retrievability first
The strongest result in the study, and the one you control outright.
- 85.9% of brands gained own-domain citations, with a median per-brand gain of 3.48×
- 1 in 8 site: queries now points at the brand's own domain
- State claims plainly on indexable pages; anything locked in a PDF, behind a form, or rendered client-side returns nothing
Add the sources that gained
Reading share moved decisively toward primary and institutional sources.
- Regulators, standards bodies, registries and professional associations all gained
- fda.gov 5.5×, nist.gov 5.0×, consumerfinance.gov 4.6× share change
- In most companies the people who maintain those records sit in compliance, not marketing
Reweight away from "best X" placements
Affiliate and consumer review media lost on both measures, and lost harder on citations.
- Editorial fell to 0.64× read share and 0.40× cited share
- Review platforms fell to 0.73× read and 0.31× cited
- The placement still works; it is worth materially less than it was in July
Prioritize comparison and consideration queries
The change concentrates on commercial intent rather than hitting every prompt equally.
- Commercial prompts went 2.12× deeper, against 1.60× for transactional
- "Best X" and "X vs Y" prompts now run the deepest research
- That is where the extra reading happens, and where the contest is
Keep the whole thing in proportion
Research effort roughly doubled. Brand visibility barely moved.
- Mention rate fell 2.9 percentage points, from 40.5% to 37.6%, and share of voice by about one, because more players get named per answer
- Six more brands cited per tracked brand, but the leader's share rose 12%
- Asked whether this helped or hurt, the honest answer is neither, yet
Frequently Asked Questions
We cannot currently separate the two. Fan-out and source lists are read from ChatGPT's research disclosure panel, so if OpenAI began showing more of a process it had always run, the source-side results would reproduce without any change in behavior. Everything measured in the answer body is unaffected either way, because inline citations are extracted from the response text rather than the panel.
Method: how we measured this
- Population. All ChatGPT responses across 5,000+ tracked brands, August 1–17, 2026, ~300k responses/day. Comparison windows are Aug 1–7 versus Aug 9–13, skipping the Aug 8 transition day and ending before the Aug 14 Reddit event.
- Fixed cohort. Site and category figures use only brands with responses on all 17 days (~131k responses/day). The raw daily brand count swings substantially from one day to the next, so without this, changes in which brands ran would masquerade as model behavior.
- Paired prompt test. The headline 11.58 → 22.54 comes from 266,182 prompt strings that ran in both windows with three or more responses each, so it's immune to prompt-mix drift. Median paired ratio 1.64×; 81.6% of individual prompts increased; 56% rose by half or more. The effect runs across the whole population.
- Share ratios are pre-normalized. Every ratio in the category tables and both charts already divides by pool growth (1.82× read, 1.68× cited), so 1.00× is neutral. Only ever compare them against 1.00×.
- Chrome excluded. About a quarter of raw citation rows are interface assets: a favicon service, image CDNs, product photos, map-widget legal links. All excluded. Including them inflates every per-answer figure, and it did on our first pass.
- Sources versus citations. Reported separately throughout, because they move differently and the distinction reverses which category wins.
- Null baselines on anything ranked. Leaderboard churn is quoted only against a same-length pre-change comparison, because the top of these lists turns over heavily by itself. Without that control the raw figure reads as a 72% reshuffle when the true effect is about 12 points. We'll hold any future rank claim in this series to the same standard.
- Paired-brand floors applied symmetrically. Concentration and variety figures use only brands clearing the volume floor in both windows. Applying the floor per window instead admitted 1,481 brands after against 1,004 before and reversed the variety result, because the later period clears a citation floor more easily by construction.
- Cross-engine control on answer length. Paired-prompt length change over the same window: ChatGPT +5.1% median, AI Overviews +4.7%, Gemini +6.9%, Copilot −1.5%, Perplexity −34.5%. ChatGPT is not distinctive, so answer length is not used as evidence of a ChatGPT-specific change. AI Mode's headline +43% was a single 12,048-character day and is excluded as noise.
- Answer body versus research panel. Inline citations are extracted from the response text, while source lists come from the research disclosure panel. Findings are labeled by which one they rest on, because only the first is immune to the disclosure question.
- Robustness. The depth increase reproduces on two independent collection providers with separate infrastructure (1.9× and 1.45×). Provider identity is not otherwise a reported dimension. The API/gateway route is ~2% of ChatGPT volume and is not used as a control.
