Key takeaways
- On Aug 8, 2026, site:-scoped queries jumped from 0.37% to 16.8% of ChatGPT fan-out queries, and average fan-out queries per response rose from about 1.08 to about 1.83 (Promptwatch Data).
- The change looks additive. Generic searches still run, and domain-scoped lookups now run next to them.
- A site: query can only return pages a search index already holds, so indexation and crawlability matter more than they did last quarter.
- The fix is topic coverage on your own domain: pricing, comparisons, alternatives, FAQs and explainers, not one thin page per topic.
- Don't panic-rewrite. Only one publisher has measured this, OpenAI has not confirmed it, and it has only been observed in ChatGPT.
What actually changed on August 8
Query fan-out is the step where an AI engine turns your prompt into several background searches before it writes an answer. You type one question. The model quietly runs a handful of differently worded searches and stitches the results together.
Until early August, those searches were almost all open-web queries. Then something flipped. According to Promptwatch Data's report on ChatGPT's site: operator shift, queries using the site: operator went from 0.37% to 16.8% of all ChatGPT fan-out queries on August 8, 2026. That is roughly 46x in a single day. Before, about 1 in 280 fan-out queries used it. After, about 1 in 6.
The share had sat at 0.3 to 0.5% for weeks, dipping to 0.15% on Aug 3 to 5, which fits a staged rollout or a pre-launch experiment.
What a site: query looks like in practice: instead of searching "best help desk software pricing", ChatGPT may search site:g2.com best help desk software or site:yourbrand.com pricing. It is asking a specific domain a question.
The number that matters more than the headline
The 46x jump grabs attention, but the more useful figure is the average number of fan-out queries per response. It rose from about 1.08 to about 1.83 on the same day and stayed there.
If site: queries had replaced generic ones, that average would have stayed flat. It didn't. So Promptwatch reads the shift as additive: your chances of appearing through ordinary searches are intact, and site: is a new, extra retrieval path on top.
To be clear about scale, this is ChatGPT's own number. Earlier industry write-ups, like 85SIXTY's piece citing AirOps data across 72,000+ AI-generated queries, describe 8 to 10 parallel queries per prompt. Those figures measure a different thing, on a different basis, earlier in the year. Don't mix them up with the ~1 to 2 per response Promptwatch sees in ChatGPT.
What we don't know
It's worth being straight about the limits before changing anything.
- Only one publisher has measured it. The only independent write-up I found, Nadia Mohamed's analysis, cites Promptwatch as its sole source and says no second measurement exists.
- OpenAI hasn't confirmed it. No announcement, changelog or documentation corroborates a rollout on that date.
- It's ChatGPT only. I found no data on whether Perplexity, Claude, Gemini or Google AI Mode adopted a similar pattern. Most fan-out literature is about Google AI Mode and a Google patent, which is a different system.
- Post-change data is thin. Fewer than two weeks when the first analyses went out.
Promptwatch's own advice is the sensible one here: don't rewrite your content strategy on the back of a platform-wide rollout. Platform-controlled search behavior can change overnight, so monitor continuously instead of auditing once. Same logic applies to the earlier ChatGPT citation drop after GPT-5.3, which also moved things without warning.
Why this still deserves attention
I'd call it an argument that was already sound, now with a bigger stake. Three things follow from how site: works.
It resolves against an index, not a live crawl
A site: query is answered by a search index. If your page sits in "crawled, currently not indexed" in Search Console, it cannot be returned by that query. You used to lose rankings for that. Now you can lose an answer slot too.
A domain-scoped question is answered by your whole domain
This is the part people miss. Ask site:yourbrand.com pricing and the engine gets back whatever your domain has on pricing. A site with one pricing page and nothing else returns one thin result. A competitor with a pricing page, an explainer, a comparison and an FAQ returns a body of material to quote from.
Old failure modes still apply, but they are scoped to you
robots.txt blocks and client-side-only rendering were already problems. Under a site: query they are now problems specifically on your domain, in the exact moment the model is looking at it.
How fan-out queries look in 2026
The site: change sits inside a broader pattern of how these sub-queries are written. Promptwatch's ChatGPT query fanout data shows a few trends:
- A single prompt can trigger 3 to 8+ searches in some cases, though the daily average is much lower.
- The average per response went from 2.15 in early December to about 1.84 by early March, then returned in April at 1.0 per response.
- Average query length fell from about 117 characters in early December to the high 80s in February and March, then to about 53 characters (6 to 8 words) from April.
Short, blunt queries. That matters for how you write headings. Third-party research adds texture: 85SIXTY reports that 95% of fan-out phrases show zero monthly search volume, that year strings appear in about 6% of them, and that words like "pricing", "free", "pros and cons" and "complaints" show up constantly. The model is checking consensus, freshness, price and risk.
What to change in your content
Here is the practical version. None of it is exotic, and most of it is worth doing even if the site: pattern fades.
1. Make sure your important pages are indexed
Start here, because nothing else matters if it fails.
- Open Google Search Console and Bing Webmaster Tools and check the page indexing report for your top 50 commercial and informational pages.
- Look specifically for "crawled, currently not indexed" and "discovered, currently not indexed".
- Confirm pages are in your XML sitemap and aren't orphaned.
- Check robots.txt is not blocking AI crawlers you want visiting (GPTBot, OAI-SearchBot, ChatGPT-User).
- Make sure the main content renders in server-delivered HTML, not only after client-side JavaScript.
A crawler like Screaming Frog is a quick way to find orphaned pages and rendering gaps.
For larger sites, log-based and enterprise crawlers such as Botify and Oncrawl show how bots actually move through your domain.
2. Treat your domain as something that gets queried
Promptwatch's recommendation is to treat your own domain as a retrieval surface. Thin category pages, broken internal search and unindexed content now cost you answers as well as rankings.
In practice, imagine someone running site:yourdomain.com plus each of these words, and ask whether the results would be good:
- pricing
- alternatives
- vs [competitor]
- integrations
- reviews or case studies
- how it works
- limitations
If any of those come back empty or thin, that is a gap.
3. Build clusters instead of one mega-page
Each fan-out query is a separate retrieval. One 5,000-word page can only match so many of them well. A set of focused pages, each owning one angle, gives the engine more things to land on and cite.
The Searchable guide to query fan-out makes the same point: cover the question cluster, not the keyword. For "best help desk software 2026", that means the pricing angle, the competitor angle and the use-case angle each get a page.
4. Write headings that read like the queries
Given the short query length (about 53 characters lately), headings that mirror how people and models actually phrase things win over clever ones. Promptwatch's example: "Best CRM for small agencies 2026" beats "What should you look for when choosing a CRM?".
A few habits that help:
- Put the entity and category at the front of titles and H2s.
- Use the angle words models add: pricing, vs, alternatives, pros and cons, limitations.
- Put a current year in titles only where the content is truly kept current. Stale years hurt.
- Answer in the first sentence under the heading, then expand.
5. Cover comparison and risk angles honestly
Fan-out loves "complaints" and "limitations". If your site never talks about drawbacks, the model gets that from Reddit or a review site instead. Write your own limitations page. Write fair comparison pages. It is slightly uncomfortable, and it works.
6. Re-map your keyword-to-heading list
Promptwatch notes that the queries generated from a prompt changed materially on Aug 8, so keyword mappings built earlier are stale. Re-run fan-out analysis for the prompts you care about, then adjust headings and page assignments. Also re-audit after any major ChatGPT update, since these behaviors shift.
For spot checks, you can also look at which sources ChatGPT actually cites. Promptwatch's content type data shows product pages at 22.4% of citations (Sep 28 to Oct 4, 2026), followed by landing pages and listicles, and the half-life report shows landing pages and docs stay cited longest while news loses half its pages within 36 days. Evergreen, domain-owned pages are the safer bet for site: lookups.
A simple before and after
Take a fictional project management tool.
Before: one "Pricing" page, a blog post titled "Why teams love us", a features page.
After: a pricing page with plan comparison, "Alternatives to [Brand]" and "[Brand] vs [Competitor]" pages, an integrations hub, an FAQ page that answers cost and cancellation questions, a limitations section, and two use-case pages. All indexed, all in the sitemap, all linked from the nav or a hub.
When ChatGPT runs site:brand.com pricing alternatives, the second site hands back material. The first hands back almost nothing.
Tools that help you track and act on this
You don't need a new tool stack to do the indexation work above. Search Console and a crawler cover most of it. To see what ChatGPT really fans out to for your prompts, and whether you get cited, a tracker helps. Here is how a few options compare for this specific job.
| Tool | Fan-out visibility | Crawler logs | Content help | Best for |
|---|---|---|---|---|
| Promptwatch | Query fan-outs per prompt, volumes, difficulty | Yes, IP-verified | Content Agent, gap analysis | Teams that want to fix gaps, not only watch them |
| Searchable | Fan-out and prompt intelligence | Limited | Content Studio | Brands wanting monitoring plus content |
| Profound | Prompt and citation tracking | Not a focus | Limited | Enterprise monitoring |
| Peec AI | Prompt tracking | No | No | Simple multi-language tracking |
| Otterly.AI | Prompt tracking | No | No | Low-cost monitoring |
| Screaming Frog | None | Log analysis via add-on | No | Indexation and rendering audits |
Promptwatch is the one I'd point to first for this topic, because it shows the actual query fan-outs behind each tracked prompt and pairs that with IP-verified AI crawler logs and a content gap analysis. That combination answers the questions that matter after August 8: which sub-queries is ChatGPT running for my prompts, which of my pages did it fetch, and where am I missing a page entirely.
If you prefer a different monitoring approach, Searchable, Profound, Peec AI and Otterly.AI all track prompts and citations, though they stop closer to monitoring.
For a wider list of options, the GEO software directory at bestgeosoftware.com compares more of them.
A 30-day plan
| Week | Task | Output |
|---|---|---|
| 1 | Indexation audit in Search Console and Bing, robots.txt and rendering checks | List of unindexed or blocked priority pages |
| 2 | Run site: style review of your own domain for 10 core topics | Gap list by topic and angle |
| 3 | Re-run fan-out analysis for tracked prompts, remap headings | Updated keyword-to-page map |
| 4 | Publish or expand the top 5 missing pages, request indexing | New cluster pages live and crawlable |
Then measure. Watch whether AI crawlers fetch the new pages, whether they get cited, and whether traffic from AI platforms shows up. Repeat monthly, because the behavior will drift.
What not to do
- Don't stuff pages with
site:phrasing. The operator is part of the query the model writes, not text your page needs to contain. - Don't generalize to other engines. Google AI Mode, Perplexity and Gemini haven't been shown to follow this pattern.
- Don't churn out thin pages to fill a cluster. A thin page returned by a
site:query is no better than no page. - Don't treat one rollout date as a permanent rule. Promptwatch's 2026 fan-out usage report and the March 2026 search updates already show how often this layer changes.
Bottom line
The site: shift doesn't call for an emergency. It strengthens advice that was already good: get your pages indexed, cover each topic from several angles, and write headings that match short, blunt queries. If you do those three things, you're protected whether ChatGPT keeps querying domains directly or quietly goes back to open-web searches next month. Start with the indexation audit this week.
If you'd rather hand the strategy and execution to specialists, an SEO and GEO agency such as 1001 SEO Media can run the technical and content side for you.