Key takeaways
- A page can rank in the top 10 on Google and still get zero citations in ChatGPT, Perplexity or AI Overviews. Ranking and citation are different selection processes.
- Google Search Console shows indexing and clicks. It has no dedicated AI citation report. Bing Webmaster Tools does: the AI Performance report (public preview since February 10, 2026) lists citations per page and sample grounding queries.
- The useful view is a join: GSC and Bing data per URL, next to AI crawler logs and AI citation data. Pages with impressions but no citations are your worklist.
- Most ranked-but-uncited pages fail for one of four reasons: bots can't fetch them, the headings don't match short fan-out queries, the content type is a poor fit, or the entity language is vague.
Why ranking and citing drift apart
For years, "rank well, get traffic" was a decent mental model. AI answers broke it. ChatGPT doesn't take your prompt and return the page that ranks first. It breaks the prompt into smaller web searches, pulls candidate pages, and cites the passages that answer each sub-question cleanly.
One secondary source, an agency write-up by Bigeye from September 2026, claims the share of AI Overview citations coming from Google's top 10 organic results fell from about 76% in mid-2025 to somewhere between 17% and 38% in early 2026. It doesn't name a study, so treat those numbers as directional. The direction matches what most SEO teams see in their own data, though: plenty of pages that sit on page one never show up in AI answers.
The good news is that this gap is measurable, and the pages inside it are the best optimization candidates you have. They already have relevance and authority in the eyes of a search engine. Something small is stopping the AI layer from using them.
What each data source actually tells you
No single tool shows the whole picture. Here is what you get from each piece.
| Source | What it shows | What it misses |
|---|---|---|
| Google Search Console | Impressions, clicks, queries, positions, indexing status | No citation-level URL counts. AI Overviews and AI Mode are folded into overall Performance reporting. Doesn't show whether AI bots can fetch a page |
| Bing Webmaster Tools AI Performance | Total citations, average cited pages, grounding queries, page-level citation counts, trend over time | Copilot, Bing AI summaries and some partner integrations only. No clicks, no ranking or prominence data, grounding queries are a sample |
| AI crawler logs | Which bots hit which URLs, status codes, errors | Doesn't show whether the page was then cited |
| AI visibility platforms | Citations across ChatGPT, Perplexity, AI Overviews, Gemini and others per prompt | Sampled prompts, so coverage depends on what you track |
| Analytics | Real sessions and conversions from AI referrers | Only measures the click, not the citation that preceded it |
The gap in GSC deserves a note. One LinkedIn post by Ran Yosef claims that when you're cited in an AI Overview and rank 7th organically, Search Console reports only your topmost position. I couldn't verify that against Google's documentation, so check it yourself before building a process around it. What is safe to say: GSC won't tell you which of your URLs got cited, and that's the whole point of pairing it with other data.
Step 1: Pull your Google Search Console baseline
Start with pages, not queries. In the Performance report, set a date range of at least 90 days, switch to the Pages tab, and export impressions, clicks, CTR and average position. If you have thousands of URLs, use the Search Console API or the Looker Studio connector to avoid the export caps.
Then build a short list: pages with solid impressions and an average position inside the top 10 for the queries you care about. Those are your "ranking" pages. Hold on to that list, because everything else gets compared against it.
A few practical notes:
- Filter by query pattern too. Questions, comparisons and "best X for Y" queries are the ones that tend to trigger AI answers.
- Segment by page type (product, blog, docs, landing). You'll need that later when you look at which formats get cited.
- Save a dated export. Citation behavior shifts week to week, and you want a baseline.
Step 2: Add Bing Webmaster Tools AI Performance
Microsoft launched the AI Performance report in Bing Webmaster Tools on February 10, 2026, as a public preview open to any verified site. Fabrice Canel of Microsoft described it as "just a preview, you will get more in 2026." Per Semrush's breakdown, it reports:
- Total citations across Copilot, AI summaries in Bing and select partner integrations
- Average cited pages (a daily average of unique URLs)
- Grounding queries, which are sample phrases the AI used to retrieve your content
- Citation counts by page URL
- A visibility trend over time
This is the only first-party, URL-level citation data you get for free from a major search company. That makes it a useful proxy even though it covers Microsoft's surfaces, not ChatGPT or Perplexity. One third-party author (Search Logistics) suggests Copilot citations may hint at ChatGPT citations given the Bing and OpenAI relationship. That's an inference, not something anyone has verified, so use it as a hint rather than a measurement.
Export the page-level table on day one. The report has no click data and can't tell a passing mention from a primary source, so you'll want your own history to see which pages are gaining or losing.
Semrush found that 39% of grounding queries were conversational questions or longer multi-word phrases. That's worth reading as a content brief: those phrases are what the AI actually searched for, and they rarely look like the head terms you rank for in GSC.
Step 3: Join the data and find the gap
Now put everything in one sheet (or BigQuery, if you're that kind of team), keyed on URL. Columns:
- GSC impressions, clicks, average position
- Bing AI Performance citations per URL
- AI crawler hits per URL (from logs)
- AI citations per URL from your tracking platform
- Page type and last updated date
Then sort into four buckets:
| Bucket | GSC | AI citations | What it usually means |
|---|---|---|---|
| Winners | Strong | Cited | Protect and refresh. Study what they have in common |
| Ranked but not cited | Strong | None | Your main worklist. Fixable with structure, access or entity work |
| Cited but not ranked | Weak | Cited | AI found a clean answer your SEO missed. Strengthen these pages for search too |
| Neither | Weak | None | Normal content triage |
Semrush suggests a similar workflow for the Bing report: cross-reference top-cited pages with top-traffic pages, mine grounding queries for content gaps, and compare indexed-but-uncited pages against cited ones. Do the same across all your sources.
The second bucket is the one that stings. Those pages already did the hard part. They have relevance, links and a position. Something about them just isn't getting picked.
Step 4: Diagnose why a ranked page isn't cited
Work through the causes in this order. The early ones are cheaper to check and more embarrassing to miss.
Can AI crawlers actually fetch it?
GSC shows indexing, not whether GPTBot, ClaudeBot or PerplexityBot can read the page. A CDN or WAF rule, a robots.txt line or a bot challenge can block AI crawlers while Googlebot sails through.
Promptwatch Data's AI crawler traffic report (IP-verified requests, June 8 to September 6, 2026) gives you a benchmark. OpenAI accounted for 79.8% of verified AI crawler requests in the week of August 31 to September 6, down from 94.8% in the week of June 8. If a provider is a large share in that data but close to zero in your own logs, you are probably blocking it. Only requests whose IP matches the operator's published ranges are counted there, so spoofed user agents don't inflate the numbers, and you should filter yours the same way.
Pull your server or CDN logs and check the status codes AI bots get on your ranked pages. A 403 or a JavaScript-only shell explains a lot. If your site renders client-side, look at a prerendering option too.
Do your headings match the queries AI actually runs?
ChatGPT doesn't search for your user's prompt verbatim. It generates its own short queries. According to Promptwatch's ChatGPT query fanout data, the average number of distinct web searches per response dropped from 2.15 in early December to about 1.84 by early March, and average fan-out query length shrank from about 117 characters to roughly 53 characters by April. Shorter queries mean your page has to match a compact phrase, not a long natural-language question.
So a page that ranks for a broad head term may not match the short sub-queries ChatGPT fires off. The fix Promptwatch's data suggests: write headings that read like short search queries, front-load the entity and category in titles and H2s, and build focused pages per sub-question (comparisons, pricing, alternatives, how-tos) instead of one mega-page.
Another data point on this: ChatGPT began using the site: operator at scale on August 8, 2026, jumping from roughly 0.4% to about 17% of fan-out queries, according to Promptwatch's site: operator report. Clear site and brand naming matters more when the model is searching inside specific domains.
Is the page the wrong content type?
Promptwatch's citation share by content type data (September 28 to October 4, 2026) shows what ChatGPT actually cites:
| Content type | Share of classified ChatGPT citations |
|---|---|
| Product pages | 22.4% |
| Landing pages | 16.8% |
| Listicles | 13.6% |
| How-tos | 8.7% |
| Documentation | 7.5% |
| General articles | 6.1% |
| Comparisons | 5.9% |
| News | 5.7% |
| FAQ | 3.3% |
| Reviews | 1.7% |
For commercial prompts, product pages climb to 26.0% and landing pages to 20.6%. If your ranked-but-uncited pages are mostly thin blog posts or news items, you may be sitting in a low-share format. News is also short-lived: per Promptwatch's content half-life report, news lost half its cited pages within 36 days, while landing pages and docs lasted longest.
Is the entity language vague?
AI engines pull passages. If a passage says "our platform helps teams do more" without naming the product, category or audience, there's nothing to cite. Check that each ranked page names the brand, the category and the specific claim in extractable sentences near the top of a section. Bigeye's write-up lists extractability, entity clarity and bot access as the usual culprits, which matches what the other data here points to.
Also check for outside corroboration. Answer engines tend to favor claims echoed across multiple places online. A page with strong rankings but no third-party mentions is a harder citation target.
Step 5: Fix, republish, and re-measure
For each page in the ranked-but-uncited bucket, pick the cheapest fix that matches the diagnosis:
- Blocked or failing for bots: change the robots.txt, CDN or WAF rule, then confirm in logs that the bot gets a 200.
- Heading mismatch: rewrite H2s as short queries, split mega-pages into focused ones, put the entity first.
- Wrong format: add a comparison table, a clear definition block, or turn a vague blog post into a landing or how-to page.
- Vague entities: name the brand, product and claim explicitly; add dated facts and sources.
- Missing corroboration: earn mentions on sources AI already cites in your category.
Then wait. Citation data moves slowly, so re-check the Bing report and your tracker after two to four weeks, not two days. Keep your baseline export so you can compare.
Tools that help with each part
You can run the free parts of this workflow with GSC, Bing Webmaster Tools and your server logs. Past that, you need something that tracks citations across engines and connects them to pages.
Search Logistics (June 2026) listed starting prices of around $29/month for Otterly.ai, $95/month for Peec AI and $99/month for Profound, with a market range up to $700+/month. That's third-party and may be stale, so verify on the vendor sites. All of them track sampled prompts.
Otterly.AI
Profound
If you want the GSC and Bing data joined with AI data in one place rather than stitched together in a spreadsheet, Promptwatch does this in its Search Insights and 360 Insights views: Search Console and Bing Webmaster traffic per page, overlaid with AI crawler visits, citations and clicks. It's built to answer exactly the question "which pages rank in Google and Bing but never get cited by AI?" It also includes IP-verified crawler logs, which cover the access check from step 4, and a Content Agent and Action Board for the fixing part.
For the Google side of the data, Semrush and Ahrefs both have AI tracking add-ons, though they work from fixed prompt sets and don't include crawler logs. For deeper log analysis on large sites, a crawler or log analyzer like JetOctopus or Screaming Frog can handle the bot access checks.
If you're comparing platforms, the GEO software directory at bestgeosoftware.com lists most of the options side by side.
Common mistakes
- Treating Bing's AI Performance report as a ChatGPT report. It covers Microsoft surfaces. Useful, but not a substitute.
- Reading grounding queries as user prompts. They are sample retrieval phrases.
- Judging after a week. Citation shifts take weeks, and ChatGPT's behavior itself changes: Promptwatch's citation drop data shows average citations per response fell around the GPT-5.3 rollout on March 4, 2026, which can look like your page lost ground when everyone did.
- Optimizing every page. Start with the ones that already have impressions.
- Skipping the baseline export.
A simple weekly routine
- Export GSC pages and Bing AI Performance page citations.
- Pull AI crawler hits and AI citations per URL.
- Refresh the four-bucket sheet.
- Pick five ranked-but-uncited pages and diagnose them in order: access, headings, format, entities.
- Ship the fixes, log the date, and re-check in a month.
Promptwatch's broader AI search research hub is a good place to check whether a shift you're seeing is yours or the whole market's before you rewrite anything.
The pages that rank but never get cited are the cheapest wins you have. Find them, check whether the bots can even read them, and match the headings to the short queries the models actually run.