Key takeaways
- Citation rank tells you where your domain lands in the source list of an AI answer. Self-citation rate tells you how often your own pages are the ones being cited when your brand is mentioned.
- Read them together. Low share of voice points to coverage, late position or poor citation rank points to prominence, and mentions without self-citations point to content or crawlability.
- Crawlability is the cheapest thing to rule out. Check robots.txt, CDN and WAF rules, and compare your own crawler logs with the provider mix in Promptwatch Data.
- Answers are crowded. 47.6% of ChatGPT responses cite 10 or more unique domains (Sep 28 to Oct 4, 2026), so plan on sharing every answer.
- Do not borrow intuition from academic bibliometrics. A high self-citation rate means something different for a journal than for a brand in an AI answer.
First, a note on the term "citation"
Search for "self-citation rate" and you land in academia. Journals, authors, h-index debates. That literature is real and interesting, and it is worth a short detour because it explains why people get confused.
In scholarly publishing, self-citation is when an author or journal cites its own earlier work. The Scholarly Kitchen put it bluntly: everyone cites themselves, and it is normal behavior for anyone who has published more than one paper. Clarivate's Journal Citation Reports analysis found that journals in the top quartile of their category have self-citation rates of 10% or less. Self-citation there is a warning signal when it is high, because it can inflate impact metrics.
The prominence side has an academic parallel too. Manuel Garcia's guide on highly cited papers notes that searching mainly through prominent citation trails can reproduce an existing imbalance, because visible work steers you toward more visible work. That is a decent mental model for AI answers: a few sources get pulled in again and again, and everyone else has to work to be seen.
But the AI search meaning is different, and the interpretation flips. In an AI answer, a high self-citation rate is usually good news. It means that when the model talks about you, it points to your own pages. A low one means it talks about you but sends readers somewhere else, or doesn't bother citing anyone at all.
The two metrics, defined for AI search
Citation rank
Citation rank is where your domain lands in the source list attached to an AI response. Position 1 is the first source shown, position 9 is buried. It is a prominence metric. It only exists when you are cited at all, so it says nothing about the prompts where you were absent.
Self-citation rate
Self-citation rate is the share of responses mentioning your brand where one of your own pages is among the cited sources. If ChatGPT names you in 100 responses and links your domain in 40 of them, your self-citation rate is 40%. The other 60 are mentions that rely on third parties, or on training data, to describe you.
Why you need both
One metric alone is misleading.
- Good citation rank with a tiny number of citations means you win rarely. That is a coverage problem hiding behind a flattering average.
- Plenty of mentions with a low self-citation rate means the model knows you but doesn't use your site. That is a content or crawlability problem.
- Strong self-citation but poor rank means you are cited, just last. That is a prominence problem.
The metrics sit alongside a few others: Visibility Score, share of voice, position in the answer, and sentiment. Tools like Promptwatch report all of these, so you can read them side by side rather than in isolation.
The diagnostic matrix
Here is the quick version. Each row is a pattern you might see in a dashboard, with the most likely cause.
| What you see | Likely problem | First thing to check |
|---|---|---|
| Low share of voice, few prompts where you appear at all | Coverage | Content gap analysis against the prompts and sub-queries AI actually runs |
| Cited, but citation rank is late (bottom of the source list) | Prominence | Page type, specificity, and third-party corroboration |
| Brand mentioned often, self-citation rate near zero | Content or crawlability | Crawler logs, robots.txt, then whether your pages answer the question |
| Self-citation high on brand prompts, absent on organic prompts | Coverage | Non-branded topic pages and third-party mentions |
| Sentiment weak even when cited | Reputation | What the cited third-party sources say about you |
| Crawler logs show a provider visiting heavily, but no citations | Content | Whether the pages fetched are actually the ones answering prompts |
| Crawler logs show a provider near zero | Crawlability | Blocks at CDN, WAF, or robots.txt |
A rule I use: always rule out crawlability first, because it is binary and cheap to check. If an AI system cannot fetch your pages, no amount of content work matters.
Start with the crowd: how many sources share an answer
Before interpreting your own numbers, you need a sense of the baseline. Promptwatch Data looked at how many unique domains ChatGPT cites per response between Sep 28 and Oct 4, 2026. The full numbers are in the unique domains per response report.
- 47.6% of responses cite 10 or more domains.
- 66.1% cite at least 8.
- Only 6% cite 3 domains or fewer, and 1% cite just one.
Broken down by prompt type, it gets more useful:
| Prompt type | Cites 10+ domains | Other figures |
|---|---|---|
| Organic (no brand named) | 52.5% | 71.8% cite 8+; 4.5% cite 3 or fewer |
| Competitor comparison | 49.1% | 66.5% cite 8+; 0.35% cite a single domain |
| Brand-specific | 21.3% | Most common count is 5 domains (14.9%); 40.7% cite 5 or fewer |
What this means for diagnosis is straightforward. On organic prompts you are one of ten or more voices, so a mediocre citation rank is normal and a self-citation rate is hard to move. On brand-specific prompts, only about five domains tend to appear, so if you are not one of them, that is a strong signal. Your own site is probably not answering pricing, features, reviews or alternatives clearly enough, and the few domains that are cited will shape how the model describes you.
So set expectations per prompt type. Expecting a 60% self-citation rate on organic prompts is unrealistic. Seeing 15% on brand prompts is a problem.
Diagnosing a coverage problem
Coverage means you are simply not in the conversation for enough prompts. The symptoms are low share of voice, a small number of prompts where you are visible, and good numbers only on branded queries.
What causes it
Often it is about the query, not the prompt. When ChatGPT searches the web it does not run the user's prompt verbatim. It generates sub-queries, called fan-outs. Per Promptwatch's query fanout data, your content needs to match one of those generated queries, not the original prompt. That report showed average fan-out length falling from roughly 117 characters in December to about 53 characters in April. Shorter, more focused queries favor focused pages.
One caution on that data: the report's April figure of 1.0 queries per response sits awkwardly next to the later finding that ChatGPT began using the site: operator at scale on August 8, 2026, with searches per response nearly doubling. Check the site: operator report before quoting current fan-out depth. In that report, site: queries jumped from about 0.4% to about 17% of fan-outs overnight.
How to fix it
- Map your pages against the prompts and sub-questions AI answers. Gaps show up quickly: comparisons, pricing, alternatives, how-tos.
- Prefer several focused pages over one mega-page.
- Write headings that read like short search queries, entity and category up front.
- Earn third-party mentions on review sites, listicles and forums, since those are where crowded organic answers draw from. Promptwatch's citation share by content type report shows product pages at 22.4% of ChatGPT citations (Sep 28 to Oct 4, 2026), followed by landing pages and listicles, so page type matters as much as topic.
Platforms that offer content gap analysis include Promptwatch, Profound and AthenaHQ, though many monitoring-first tools stop at showing the gap rather than helping you fill it.
Diagnosing a prominence problem
Prominence means you are cited, but low in the source list, or in the answer text late. Your coverage is fine. Your weight is not.
Reading the signals
- Citation rank is consistently in the bottom half.
- Position in answer is late even when you are mentioned.
- Your cited pages are thin, generic or older than competitors' pages.
The academic analogy helps here. Garcia's point about highly cited papers is that prominence compounds: visible sources get followed, which makes them more visible. AI answers behave similarly when a handful of domains dominate a topic. The consolation is that dominance is not purely about authority. Promptwatch's domain rank report found that sites with domain rank 46-75 earned almost half of ChatGPT citations in August 2026, while DR 91-100 sites fell to about 3%. Mid-authority sites with the right page win a lot.
How to fix it
- Make the page the best answer to one specific question rather than a general overview.
- Put the direct answer near the top, with the entity and category named plainly.
- Refresh stale pages. ChatGPT content half-life data shows landing pages and docs stay cited longest while news loses half its pages within 36 days, so match your refresh cadence to the content type.
- Build corroboration on third-party pages that AI already cites.
Track where you appear in the response and how you are described, not just whether you are cited. That is also the advice in the unique-domains report.
Diagnosing a crawlability problem
This is the one with the clearest signature: mentions without self-citations. The model knows your brand, probably from training data or from other sites, but your own pages are not being retrieved.
Check the plumbing first
- Open robots.txt and look for blanket disallows or rules aimed at AI user agents.
- Check CDN and WAF rules. Bot protection often blocks legitimate AI fetchers without anyone noticing.
- Verify that key pages return a clean 200 to AI crawlers, not a challenge page or a JavaScript-only shell.
- Look at your server or CDN logs per provider.
Promptwatch Data's AI crawler traffic report is useful as a benchmark. It uses IP-verified requests, with spoofed user agents excluded, covering June 8 to September 6, 2026. OpenAI made up 79.8% of verified AI crawler requests in Aug 31 to Sep 6, down from 94.8% in Jun 8 to 14. The practical test from the report: if a provider is large in the chart but near zero in your logs, you are likely blocked or unreachable for that provider.
Provider totals also hide the split between training crawlers and answer-time fetchers. A site can be crawled heavily for training and still blocked from the user-triggered fetcher that produces citations. Switch to the per-crawler view.
Tools that help
| Need | Options |
|---|---|
| AI crawler logs and crawl-to-citation path | Promptwatch, DarkVisitors |
| Traditional technical crawl of your site | Screaming Frog, Sitebulb, JetOctopus |
| Log-file analysis at enterprise scale | Botify, Oncrawl |
| llms.txt generation and validation | llmtxt.info |
A generic crawler shows what a search engine can see. It will not tell you what GPTBot, OAI-SearchBot or ClaudeBot actually requested, and that is the question here. For the AI side you want logs that separate verified AI bots from spoofed ones. Promptwatch's crawler analytics works with Cloudflare, Vercel, Fastly and raw server logs, and also surfaces plain-language reasons when a crawl fails.
When crawlability is fine but self-citation is still low
Then it is a content problem. The bot arrives, reads the page, and cites someone else. Common reasons:
- The page does not directly answer the prompt. It talks around it.
- Facts are buried, vague, or contradicted elsewhere on the site.
- A third-party page says it better, with specifics.
Compare crawls to citations per page. A page that is crawled often but never cited is a rewrite candidate. A page never crawled is a discovery or access problem.
A worked example
Say you run a mid-sized project management tool and you track 100 prompts for a month.
- Brand-specific prompts (20): you are mentioned in 19, self-cited in 6. Self-citation rate is about 32%.
- Organic prompts (60): you are mentioned in 9. Share of voice is low.
- Competitor comparison prompts (20): you are mentioned in 14, self-cited in 3.
Reading it:
- Organic: mentioned in 9 of 60 is a coverage problem. You are missing from the broad category conversation.
- Brand-specific: nearly always mentioned but rarely self-cited. With only about five domains in a typical brand answer, being absent from the source list matters. Check crawler logs for those pricing and feature pages first. If they are being fetched, rewrite them to answer pricing, features, reviews and alternatives plainly.
- Competitor comparisons: low self-citation suggests third-party comparison pages are winning. You need your own comparison content and presence on the listicles that get cited.
Three different problems, three different workstreams, one dataset. That is the value of reading the two metrics together.
Measuring over time, not once
One prompt tells you almost nothing. AI answers vary run to run, models update, and citation behavior shifts. Promptwatch Data recorded reddit.com's share of ChatGPT citations dropping from about 4% to about 0.5% on August 14, 2026, and average citations per response fell around the GPT-5.3 rollout on March 4, 2026 (see the citation drop report). Your metrics can move for reasons that have nothing to do with your work.
So:
- Track across many prompts, grouped by type.
- Compare time periods, not single days.
- Annotate your own changes (page rewrites, robots.txt edits) so you can attribute movement.
- Watch the platform mix. Social citations, for instance, are very uneven: ChatGPT drew 0.46% of citations from social platforms while Google AI Overviews drew 12.2% (Sep 28 to Oct 4, 2026, social media citations by AI model). A self-citation target that works for one model may be wrong for another.
Among tools that track citations and rank per model, options include Promptwatch, Peec AI, Otterly.AI and Semrush One. Monitoring-only trackers can show you the numbers. Fewer connect them to crawler logs and traffic, which is where the diagnosis usually gets settled. For a wider view, the GEO software directory at bestgeosoftware.com lists more options.
A practical order of operations
- Rule out crawlability. Check robots.txt, CDN rules and per-provider logs.
- Segment prompts by type: organic, brand-specific, competitor comparison.
- Read self-citation rate on brand prompts. This is your most controllable number.
- Read share of voice on organic prompts. This is your coverage number.
- Read citation rank and position on prompts where you appear. This is your prominence number.
- Fix the biggest gap, ship, then re-measure over several weeks.
Do not average across prompt types. A blended self-citation rate hides the exact problem you are trying to find.
Common mistakes
- Treating self-citation like the academic metric. Higher is generally better here, not suspicious.
- Chasing citation rank on prompts where you are barely present. Fix coverage first.
- Assuming a block is a content problem. Always check logs before rewriting.
- Judging on one prompt or one day.
- Expecting identical behavior across models. Real interface behavior differs from API output, so track the user-facing surface where possible.
The short version
Citation rank and self-citation rate answer different questions. Rank says how prominent you are when cited. Self-citation says whether your own site is the one doing the work. Put them next to share of voice and crawler logs, and the problem usually announces itself: missing from answers is coverage, low in the list is prominence, named but not linked is crawlability or content. Start with the logs, because that is the fastest thing to eliminate.