Two named datasets disagree. Promptwatch measured ChatGPT site:-scoped fan-out queries jumping from about 0.37% to 16.8% on 8 August 2026. DEJAN AI measured 0.14% across 196,692 OpenAI fan-out queries that month, and 0 of 371 on 8 August itself. Our own 40-poll ChatGPT series across the same window shows no step change either.
Most coverage of this question reports one of those numbers as settled and never mentions the other. This page reads both datasets at the page of the organisation that published them, adds a third from our own poll registry, and says where the three disagree, why they might, and what test would settle it.
What is actually claimed to have changed on 8 August 2026?
The claim has two halves, and they are separate measurements that get quoted as one thing.
Half one, the fan-out claim. Promptwatch states that the share of ChatGPT fan-out queries containing the site: operator — which scopes a search to one named domain — “jumped from about 0.37% to 16.8% of all fanout queries in a single day, roughly a 46x increase”, and that “the average number of searches ChatGPT runs per response nearly doubled, from about 1.08 to 1.83”. Both figures appear in Promptwatch’s own write-up, bylined Klaas Foppen, whose page metadata gives a publication date of 20 August 2026 and a modified date of 8 September 2026 (read 16 September 2026). DEJAN AI describes the same write-up as published on 10 August 2026; the byline and the page metadata both say 20 August, so that is the date we use.
Half two, the Reddit claim. Promptwatch separately reports that reddit.com held an average 3.83% of ChatGPT Search citations from 18 July to 7 August 2026, fell below 1% on 14 August, and averaged 0.52% from 14 to 17 August — an 86.4% relative drop. Promptwatch flags that number itself: on its own data report it says a data-collection issue cannot be ruled out, and asks readers to treat the size of the drop as provisional (read 16 September 2026). Google AI Overviews moved 2.37% to 2.10% over the same window and Google AI Mode 2.22% to 1.54%, both gradually.
Note the dates. The fan-out change is dated 8 August; the Reddit collapse is dated 14 August. Promptwatch proposes the first as the mechanism for the second. That proposed link, not either measurement, is the load-bearing part of the story everyone repeated.
What is NOT in dispute, and who disputes what
This is the boundary, and getting it wrong is how the argument goes in circles.
DEJAN AI, in a post by Dan Petrovic dated 25 August 2026, disputes only the fan-out measurement. It captured 551,695 fan-out queries between 1 March and 25 August 2026 across providers; the OpenAI subset is 196,692 fan-out queries from three GPT-5 versions, called through the OpenAI Responses API with web search enabled, driven by a fixed monitoring panel run daily. Its August site: share is 0.14%, its highest single day in August is 0.86%, and on 8 August itself “0 of 371 fanouts contained site:“. Its fan-out-per-response figure is 1.09 in July and 1.10 in August — flat, and close to the 1.08 Promptwatch reports as the pre-change value, nowhere near 1.83.
DEJAN does not address Reddit citation share at all. As at 16 September 2026, nothing in that post contests the 3.83% to 0.52% figure. The disputed claim is therefore why Reddit citations fell in ChatGPT, not whether they did.
Promptwatch is itself more careful than its resellers. Its own post says exact percentages “will vary by methodology (which brands are sampled, which dates are compared, how citation is defined), so treat the direction and the timing as the reliable part, not any single tool’s decimal point”. It also refers to “other AI-visibility trackers” reporting a comparable drop, but as at 16 September 2026 no such tracker is named or linked anywhere on that page, so we have not counted it as corroboration.
Three datasets, three surfaces: the methodology table
The single most useful thing you can do with a disputed measurement is line the methods up rather than the headlines. This is the same test we set out in our guide to reading a 2026 AEO benchmark report: whose panel, which months, what was actually counted.
| Dimension | Promptwatch | DEJAN AI | Zian AI (this page) |
|---|---|---|---|
| Surface sampled | ChatGPT consumer interface | OpenAI Responses API, web search enabled | OpenAI Responses API, web search enabled |
| Window | 18 Jul to 17 Aug 2026 (citations); 8 Aug step (fan-out) | 1 Mar to 25 Aug 2026 | 10 Jul to 14 Sep 2026 |
| Sample size | Not stated on the post | 196,692 OpenAI fan-out queries (551,695 all providers) | 573 scored answers across 40 polls |
| What was counted | Fan-out query strings; domain share of citations | Fan-out query strings | Citation URLs attached to the delivered answer |
| Model version disclosed | No | Yes: gpt-5.2, gpt-5.4, gpt-5.5 (gpt-5.5 carries all August traffic) | Partially: the gpt-5 alias, snapshot not pinned |
site: share, August |
16.8% on 8 Aug, climbing | 0.14%; 0 of 371 on 8 Aug | Not measured — we do not capture fan-out queries |
| Fan-outs per response | 1.08 to 1.83 | 1.09 (Jul) to 1.10 (Aug) | Not measured |
Read the bottom two rows honestly. Our poll registry cannot arbitrate the fan-out claim, because we do not record fan-out queries. What it can do is test the downstream observable: if ChatGPT started issuing roughly 70% more searches per response, the number of sources that ended up cited in the answer is where you would expect to see it.
What our own prompt registry recorded across the disputed window
Method and window, stated in full. Zian AI runs a fixed prompt registry — brand, category and vertical questions — polled against two engines every autopilot run, with every raw answer and every cited URL written to disk. The registry held 14 prompts from 10 July to 1 September 2026 and 16 from 2 September, when two head-to-head comparison prompts were added, so all eight of the 16-prompt polls fall on the later side of the split below; the table therefore also reports the like-for-like figure restricted to the 14 prompts that ran throughout. The ChatGPT leg calls the OpenAI Responses API with the web_search_preview tool, model alias gpt-5, reasoning effort low. Between 10 July and 14 September 2026 that produced 40 ChatGPT polls and 573 scored answers. “Cited” means zian.ai appeared as a formal URL citation on the response, or, failing that, the brand was named in the answer prose; 41 of our 43 ChatGPT citation events were formal URL citations and 2 were prose-only. The counts below are of citation annotations as the API returns them, so a source cited twice inside one answer counts twice and these are not distinct-domain counts. The citation list is truncated at 25 URLs per answer, so the means below are censored at 25.
| Measure | 10 Jul to 7 Aug 2026 | 8 Aug to 14 Sep 2026 |
|---|---|---|
| Polls | 18 | 22 |
| Answers scored | 249 | 324 |
| Mean citation URLs per answer | 15.68 | 15.91 |
| Mean, restricted to the 14 prompts that ran throughout | 15.68 | 15.89 |
| Median citation URLs per answer | 16 | 16 |
| Answers returning zero citations | 6 (2.4%) | 12 (3.7%) |
| Answers at the 25-URL cap | 15 (6.0%) | 17 (5.2%) |
| Polls citing zian.ai at least once | 18 of 18 | 22 of 22 |
This is a flat series and we are reporting it as flat. The 8 August poll itself returned a mean of 16.93 citations per answer, which is inside the pre-window range of 13.21 to 17.00 and below the highest pre-window poll. The Gemini leg, on gemini-2.5-flash with Google Search grounding, moved the other way and only slightly: 8.88 citations per answer before 8 August, 8.10 after. Neither engine shows a discontinuity on the disputed date.
The honest caveat is the one that matters most: our ChatGPT leg is an API surface, the same class of surface DEJAN sampled. It is a second API dataset, not an independent check on the consumer interface Promptwatch samples. It corroborates DEJAN and it does not test Promptwatch. We also do not pin a model snapshot, so we cannot attribute our flatness to a particular version. Sixteen prompts on one domain is a small panel, and we designed it that way; what to store and what not to trust in a panel like this is set out in our guide to what an AEO poll harness should record.
The conditions under which the answer changes
Four variables could each produce this gap on their own, and only one of them is anybody being wrong.
Model version. This is DEJAN’s strongest finding and the one the agency reposts drop. Splitting its OpenAI capture by version puts the change at the version boundary rather than at any date in August: gpt-5.2 scoped 5.88% of fan-outs to a domain, gpt-5.4 5.32%, and gpt-5.5 just 0.15% — “one in 647”. DEJAN states that gpt-5.5 carried all of its August traffic. Now add a fact neither party states. OpenAI’s own API changelog (read 16 September 2026) dates the GPT-5.5 release to 24 April 2026 and the GPT-5.6 family release to 9 July 2026. DEJAN’s August panel was therefore running a model that had been superseded on the API a month before the disputed date, while the consumer interface Promptwatch samples was running whatever OpenAI was serving. If site: use tracks version, as DEJAN’s own table says it does, that alone can produce two datasets this far apart with neither being wrong.
Prompt mode. DEJAN splits its runs into short repeated visibility prompts and longer ad-hoc citation-mining prompts, and reports 6.29% site: use for citation mining against 0.44% for visibility tracking across the full window, “7x to 91x” apart in every month where both ran. Fixed monitoring panels — ours included — sit on the low side of that split by construction.
Reasoning mode. This is the candidate mechanism neither post raises, and it matters because it predicts both Promptwatch observations at once. A Semrush study run with Kevin Indig, published 30 June 2026 and read 16 September 2026, found that switching the same 100 prompts from minimal to high reasoning took web searches from 245 to 1,130, citations per response from 2.6 to 4.5, and Reddit citation share from 15% down to 7%. We covered it in full in how ChatGPT Thinking mode cites more and different sites, so we will not re-run the numbers here. The point for this page: more searches per response plus a falling Reddit share is exactly the signature a routing shift towards deliberative answers would leave, with no change to site: behaviour required. Our own poll runs at low reasoning effort, which means our flat series is consistent with a reasoning-mode explanation rather than evidence against one.
Surface. DEJAN names this one itself: “Promptwatch samples the ChatGPT consumer interface. We sample the OpenAI API with web search.” Prompt wording and prompt mix differ too, and DEJAN says plainly that “our data does not separate them”.
What would settle it, and what to check on my own site
DEJAN names the test and states that it has not been run: “Running our own prompt panel through the ChatGPT interface on the same day as the API run would hold the prompts constant and leave the surface as the only variable. We have not run that test.” Until somebody does, the correct description of the August 2026 change is unresolved on the consumer surface, absent on the API surface, and confounded by at least three variables. A dated first-party statement from OpenAI would also close it, and there is not one: its API changelog carries no entry dated 8 August 2026 at all — its August entries are the 4th, 5th, 6th, 7th, 13th, 20th, 21st, 26th and 29th, none concerning search behaviour. That is evidence about the API only, and we could not check the consumer release notes because help.openai.com returned HTTP 403 to every route we tried on 16 September 2026.
Here is our decision rule for readers with a site to run, rather than an argument to win. We call it the two-surface rule: act on a retrieval claim only when you can see it on your own logs or your own panel, and only when it has held for three consecutive observations.
| What you can observe | Do this | Do not do this |
|---|---|---|
| Your own ChatGPT citations fell across many Reddit threads, three polls running | Rebalance effort towards pages on your own domain that answer the prompt directly | Conclude the cause; the mechanism is still disputed |
| Loss concentrated in one or two subreddits or topics | Treat it as topic-level churn and investigate those threads | Attribute it to a platform-wide retrieval change |
| Two observations, not three | Keep polling | Rewrite your content plan; single-poll movement is noise |
| OAI-SearchBot or GPTBot hits on your domain changed after 8 August in your server logs | Report it with the raw counts and the dates; this is first-party evidence | Report it without splitting by response status code |
| You have no per-engine measurement at all | Stand up a fixed panel of 12 to 40 prompts before changing anything | Act on any vendor chart, including the ones on this page |
Standing that panel up yourself is doable and we would encourage it. The honest cost: a fixed prompt list you never edit in place, two API keys, a scheduler, somewhere to store every raw answer rather than the score, and a person who notices when an engine changes its response schema and silently zeroes your counts. Ours is roughly 40 lines of polling code per engine and it still needed hand-holding through a Cloudflare error class in July. The part that breaks at volume is not the polling, it is the discipline: retiring prompt IDs instead of editing them, and refusing to publish a number that moved once. Whether that is worth a maintainer is arithmetic only you can do.
If a domain-scoped or deliberative retrieval step is now reading your site directly, the page it reads has to answer the buyer question on its own, with sources — which is why our category explainers, such as the one on what autonomous AI sales agents actually are, are written as documentation rather than brochures. Two adjacent questions have their own data and their own pages: the engine-mix picture is in ChatGPT’s shrinking share of AI referrals, and how much of the citation pool user-generated content holds is in Reddit, YouTube and the UGC share of AI citations.
Frequently asked questions
Did ChatGPT change how it searches in August 2026?
Unresolved, as at 16 September 2026. Promptwatch measured the share of ChatGPT fan-out queries using the site: operator rising from about 0.37% to 16.8% on 8 August 2026 on the consumer interface. DEJAN AI measured 0.14% for August across 196,692 OpenAI fan-out queries on the API, with 0 of 371 fan-outs on 8 August containing the operator. Both are first-party captures of different surfaces.
Does ChatGPT still cite Reddit?
Less than it did, on the one dataset that measured it. Promptwatch reports reddit.com averaging 3.83% of ChatGPT Search citations from 18 July to 7 August 2026 and 0.52% from 14 to 17 August, a relative fall of 86.4%. Promptwatch flags the size of that drop as provisional on its own data report, saying a data collection issue cannot be ruled out. That figure is not contested by the DEJAN AI dataset, which measures fan-out queries rather than citation share and does not address Reddit.
Why do the two datasets disagree?
Surface, prompt mix and model version. DEJAN AI sampled the OpenAI Responses API and Promptwatch sampled the consumer interface. DEJAN also reports that site: use tracks the model version: gpt-5.2 at 5.88 percent, gpt-5.4 at 5.32 percent and gpt-5.5 at 0.15 percent, with gpt-5.5 carrying all of its August traffic. The OpenAI API changelog dates the GPT-5.6 family release to 9 July 2026, a month before the disputed date.
Did OpenAI announce a search change on 8 August 2026?
Not on the API changelog. We read the OpenAI API changelog on 16 September 2026: it carries no entry dated 8 August 2026, and none of its August entries concerns web search behaviour or query fan-out. That is evidence about the API surface only. The ChatGPT consumer help centre returned HTTP 403 to every route we tried on the same date, so we could not check that surface and we are not claiming anything about it.
What did the Zian AI poll registry show across the window?
Nothing moved. Across 40 ChatGPT polls of a fixed prompt registry between 10 July and 14 September 2026, the mean number of citation URLs per answer was 15.68 before 8 August across 249 answers and 15.91 after across 324 answers, with a median of 16 on both sides. Restricted to the 14 prompts that ran across the whole window the later mean is 15.89. The list is truncated at 25 URLs per answer. This is one site, a small panel, and an API surface rather than the consumer interface, so it corroborates the API reading and does not test the consumer one.
Should I stop investing in Reddit for AI visibility?
Not on this evidence alone. The measured fall is specific to ChatGPT, the mechanism is disputed, and Google AI Overviews moved only from 2.37 percent to 2.10 percent over the same window on the same Promptwatch data. The defensible response is to measure your own citations per engine, split rather than blended, and to require three consecutive observations before changing a content plan.
Zian AI builds autonomous AI sales agents that work the pipeline your content earns — live phone, SMS, email and WhatsApp outreach in 30+ languages, with SmartReach AI™ orchestrating message, channel and timing, and PrecisionPitch AI™ split-testing scripts against real outcomes. Apply For Partnership.