Key takeaways
- AI engines like ChatGPT turn one prompt into several background searches called fan-out queries. Your page gets cited for those sub-queries, not for the prompt a human typed.
- A coverage score is simple: list the sub-queries, mark each one as fully, partially, or not answered by your page, and divide the points by the maximum.
- Fan-out behavior changes fast. ChatGPT's use of the site: operator jumped about 46x in a single day in August 2026, so re-run your score regularly and use current sub-queries, not a frozen list.
- Fix gaps with focused sections or separate pages, with headings that read like search queries.
- You can do all of this in a spreadsheet. Dedicated tools just save you the collection step.
Why a page can rank and still miss the answer
Here is a pattern I see constantly. A team picks a prompt, say "best CRM for small agencies". They track it, write a page to answer it, and wait. The page never gets cited. They assume the content is weak.
Often the content is fine. The problem is that the AI never searched for that prompt. It searched for something like "CRM pricing agencies 2026", "HubSpot vs Pipedrive", and "site:g2.com CRM small agency". Your page answered the question a human asked. It did not answer the questions the machine asked.
That process is query fan-out: the engine rewrites a prompt into several sub-queries, runs them in parallel, and merges the results into one answer. Searchable, Conductor, Similarweb and Ahrefs have all written explainers on it, and they agree on the mechanics. What is less well covered is a dead-simple way to check your own page against it. That is what this guide gives you.
What the data says about fan-outs
Before building a score, it helps to know what you are scoring against. Sub-queries are short and they change.
Promptwatch's ChatGPT query fanouts report tracks this over time. Some of what it shows:
- The average was 2.15 distinct searches per response in early December 2025, easing to about 1.84 by early March 2026. The report's headline says a prompt can trigger "3-8+" searches, but its own time series sits lower than that.
- The average query length shrank from roughly 117 characters in December to about 53 once fanouts returned in April 2026. That is around six to eight words.
- Short queries mean entity-plus-category phrasing: a product, a category, a modifier like a year or "pricing".
Then in August things moved again. Per Promptwatch Data on the site: operator, on August 8, 2026 the share of ChatGPT fanouts using site: jumped from 0.37% to 16.8%, about 1 in 6, and average fanout queries per response rose from about 1.08 to about 1.83 the same day. Searchable separately reports that answers running multi-query research rose from 2% to 57% in August 2026. That is a third-party figure, but it points the same direction.
The lesson for scoring: a sub-query list is a snapshot. A list you built in July is stale. Treat the score as something you re-run, not something you do once.
The scoring method in five steps
The method has one principle: score the page against sub-queries, not against the prompt.
Step 1: Pick one page and one prompt
Start narrow. One URL, one target prompt. For example, a page titled "Best project management software for agencies" and the prompt "best project management software for small agencies".
Step 2: Collect the sub-queries
This is the hard part, and it varies by engine.
- ChatGPT and Perplexity expose the searches they run. You can see them in the interface or through tooling.
- Google does not show which searches it runs for AI Overviews or AI Mode, per Searchable. You have to infer them.
Practical ways to get a list of 10 to 20 sub-queries:
- Run the prompt in ChatGPT with search on and read the searches it performs. The free GEO Inspector Chrome extension from Promptwatch shows fan-outs and fetched versus cited sources for ChatGPT.
- Use a fan-out generator to simulate likely sub-queries. Wellows and LLMClicks both offer free ones, and Promptwatch has a ChatGPT Query Fan-Out Generator.
- Pull People Also Ask questions and related searches for the prompt.
- Check Search Console for queries your page already gets impressions for.
Deduplicate. Merge near-identical phrasings. Keep the list to 10 to 20.
Step 3: Tag each sub-query by type
Tagging matters because gaps cluster by type. Common ones:
- Comparison (X vs Y)
- Pricing or cost
- Alternatives
- How-to or setup
- Best-of or list
- Modifier (year, location, audience, budget)
- Site-targeted (
site:g2.com,site:reddit.com)
Step 4: Score each sub-query 0, 1, or 2
Read your page and ask whether it answers this sub-query on its own. Be strict.
| Score | Meaning | Test |
|---|---|---|
| 2 | Fully answered | A heading or section matches the sub-query, and the answer is specific (numbers, names, steps) |
| 1 | Partially answered | The topic is mentioned but vaguely, buried, or without specifics |
| 0 | Not answered | Nothing on the page addresses it |
Similarweb's tooling uses the same full, partial, or none logic, which is a good sign the scale is reasonable. Resist the urge to give 2s generously. If a human would have to scroll and infer, it is a 1.
Step 5: Calculate and read the score
Coverage score = total points / (number of sub-queries x 2) x 100
A worked example with 12 sub-queries:
| Sub-query | Type | Score |
|---|---|---|
| best project management software agencies 2026 | Best-of | 2 |
| asana vs monday agencies | Comparison | 1 |
| project management software pricing agencies | Pricing | 0 |
| clickup alternatives for agencies | Alternatives | 0 |
| project management software client approvals | How-to | 1 |
| site:g2.com project management agencies | Site-targeted | 0 |
| free project management tools small team | Modifier | 2 |
| how to set up agency project workflow | How-to | 1 |
| project management software time tracking | Feature | 2 |
| trello vs asana | Comparison | 0 |
| project management software reviews | Best-of | 1 |
| project management software for 10 person agency | Modifier | 0 |
Total points: 10. Maximum: 24. Score: about 42%.
As a rough reading (my own rule of thumb, not an industry standard):
- 75% or higher: strong. Look for freshness and citation issues instead.
- 50 to 74%: workable. Fill the zeros first.
- Below 50%: the page answers the prompt but not what the engine searches for.
The type column is where the insight lives. In the example, all the comparison and pricing sub-queries are weak. That tells you exactly what to add.
Weighting: when raw coverage misleads
A flat average treats all sub-queries equally. They are not. Two adjustments are worth making if you want a sharper number.
Weight by recurrence. If a sub-query appears across many runs of the prompt, it matters more. Searchable's query fanout view records occurrences per sub-query, so you can weight by that. Without a tool, run the prompt five or so times and count repeats yourself.
Weight by intent. For a commercial page, comparison and pricing sub-queries deserve a 2x weight. For an informational page, how-to sub-queries do.
Keep it simple though. If the weighted score and the flat score tell the same story, use the flat one.
Beyond coverage: score what happens when you are absent
Coverage on the page is only half the check. Searchable suggests measuring three things: the sub-queries your brand appears for, your share of voice against competitors, and the sources engines cite on sub-queries where you do not appear.
That third one is underrated. For every zero in your table, look at which pages the engine cited instead. You will usually see one of three things:
- A competitor page with a dedicated section or URL for that exact angle.
- A third-party review or listicle (G2, a comparison blog).
- A forum or community thread.
That tells you whether the fix is a new section, a new page, or an off-site mention. It is also worth knowing what content types win. Promptwatch's data on ChatGPT citation share by content type shows product pages as the most cited type at 22.4% of citations (Sep 28 to Oct 4, 2026), followed by landing pages and listicles. If your zero-scoring sub-query is a pricing question, a clear product or pricing page may beat a paragraph in a blog post.
And be aware of how many sources are in play. Almost half of ChatGPT responses cite 10 or more unique domains in the same period, so you are rarely competing for a single slot. Covering a sub-query well gets you into a crowded set, not a monopoly.
Fixing the gaps
Once you have a scored table, the fixes are fairly mechanical.
Add sections with query-shaped headings
Promptwatch's fanout report recommends headings that read like search queries ("Best CRM for small agencies 2026") rather than conversational ones, with the entity and category at the front of titles and H2s. Queries are around six to eight words, so a heading like "Asana vs Monday for agencies" matches better than "So how do these two stack up?"
Split when the topic is too big
If you have five zeros in different types, one page will not hold them. Promptwatch's recommendation is a cluster of focused pages instead of one mega-page. A dedicated alternatives page, a pricing page, and a comparison page each match a sub-query type cleanly.
Treat your own domain as a retrieval surface
Because of the site: shift, ChatGPT may now run something like site:yourdomain.com [topic]. If your key pages are not crawlable or indexed, or your category pages are thin, those searches return nothing. Check indexation before you write a word. A page cannot score for a sub-query it never gets retrieved for.
Add structure the engine can lift
Short definitions, tables for comparisons, explicit prices, and schema markup all make an answer easy to extract. Conductor also recommends topic clusters and structured data as part of fan-out optimization.
Tools that help with the collection step
You can do everything above in a spreadsheet. Tools mainly speed up Step 2 and the re-checking.
| Tool | What it helps with | Notes |
|---|---|---|
| Promptwatch | Query fan-outs per prompt, prompt volumes and difficulty, content gap analysis, citation tracking | Tracks real UIs of ChatGPT, AI Overviews, AI Mode and more; free GEO Inspector extension shows ChatGPT fan-outs |
| Searchable | Query Fanout view lists every sub-query per prompt, with occurrences and a Track button | Good for turning sub-queries into tracked prompts |
| LLMClicks | Fan-out generator and coverage analyzer that simulates 20+ sub-queries | Quick check of whether content covers them |
| Similarweb | Scores content as fully, partially, or not answering each sub-query | Part of its AI search tooling |
| Conductor | Coverage measurement within an AEO visibility platform | Enterprise-oriented |
| Google Search Console | Real queries your page already earns impressions for | Free, and useful as a sanity check on guessed sub-queries |
Promptwatch is the one I would reach for if you want the loop closed, since it shows sub-queries, flags content gaps, and then tracks whether the page you fixed actually gets crawled and cited.
If you want to compare more options, the GEO software directory at bestgeosoftware.com lists tools in this category.
A reusable template
Copy this into a sheet. One tab per page.
| Column | What goes in it |
|---|---|
| Sub-query | The exact phrase the engine searched |
| Engine | ChatGPT, Perplexity, inferred for Google |
| Type | Comparison, pricing, alternatives, how-to, modifier, site-targeted |
| Occurrences | How many runs it appeared in |
| Score | 0, 1, or 2 |
| Page section | The heading that answers it, or blank |
| Who is cited instead | URLs of pages the engine used |
| Fix | Add section, new page, off-site mention, technical |
| Date checked | So you know when it goes stale |
Add a summary cell at the top with the score and the date. Over a few months you get a trend line, which is more useful than any single number.
Common mistakes
- Scoring against the prompt. If your list is just paraphrases of the prompt, you are measuring nothing new. Use actual sub-queries.
- Generous scoring. A passing mention is a 1 at best.
- Freezing the list. Behavior changed in August 2026 and it will change again. Promptwatch's advice is to re-run fanout analyses, since keyword mappings built before August 8 are stale.
- Chasing 100%. Some sub-queries are noise. A page that tries to answer 20 loosely related questions becomes unreadable. Cover what recurs and matters.
- Ignoring Google's opacity. For AI Overviews and AI Mode, your list is an informed guess. Label it that way.
How often to re-run it
A reasonable cadence: monthly for your top 10 revenue pages, quarterly for the rest, and immediately after a notable platform change. Also re-run after you publish a fix, to confirm the score moved and that the page is being retrieved at all.
Bottom line
A fan-out coverage score is a spreadsheet exercise: get the sub-queries, score each 0 to 2 against your page, divide by the maximum, and read the gaps by type. The number matters less than the zeros, because each zero is a specific section or page you are missing. Do it for one page this week, and you will probably find the thing you thought you had covered is the thing the engine is actually searching for elsewhere.