Citation Gaps: Finding Who AI Quotes Instead of You

A citation gap is the distance between the prompts that make you money and the URLs answer products actually quote. If ChatGPT Search answers a buyer prompt with three competitor pages and a directory, that list is your content and PR brief — not a mystery and not a share-of-voice score you invented from one screenshot.

This spoke is the competitive layer inside the Answer Engine Optimization playbook. Pair it with measuring AI search visibility so the log and the KPI sheet stay the same object. I have been SEO certified since 2021. The AEO version of that work is still a dated panel, not a vendor percentage you cannot audit.

The short answer

• Freeze 25–40 buyer prompts and a 3–6 competitor set before you change a page
• Run the panel in ChatGPT Search, Perplexity, and one Overview or AI Mode check where the query triggers
• Log brands named, URLs cited, and your status as four separate facts
• Every absent money prompt becomes a ticket: page, PR, listing, entity, or wontfix
• Re-run the same panel. One session is an anecdote; a month of the same prompts is a brief

What is a citation gap, exactly?

A citation gap is a prompt where an answer product names or links someone else — and you are absent, named without a URL, or described wrong. Semrush splits the same idea into mention, prompt, source, citation, and narrative gaps in its AI visibility-gap guide. Keep those rows separate. Collapsing them into one “AI SOV” number is how teams fund the wrong page.

Gap	What you logged	What it is not
Mention gap	Brand never appears in the answer text	A ranking drop on a blue link
Citation gap	Your domain is not in the Sources / footnotes	A missing backlink report
Source gap	A third-party URL is cited and never mentions you	“We need more blog posts”
Narrative gap	You appear, but the facts or framing are wrong	A sentiment widget with no URL
Scope gap	The prompt is outside the offer you actually sell	A crisis; it is a wontfix

Semrush’s metric glossary defines a mention as a prompt where the brand appears in the answer, a citation as a response that credits your domain, and share of voice as a percentage of mentions versus a chosen competitor set. Use those definitions if you buy the tool. Do not invent a percentage from one chat window.

Pew’s June 2026 Americans and AI study (fielded Feb 17–23, 2026; 5,119 U.S. adults) found about half of U.S. adults use AI chatbots and six in ten read AI summaries at the top of search. Those surfaces already have a source list. Your job is to read it.

How do I see what ChatGPT cites?

Use the product the way a buyer does, then write down the Sources panel. OpenAI’s ChatGPT Search help is the operator check: when search ran, you may get inline citations you can hover, and a Sources control under the response that opens cited links plus other relevant URLs. OpenAI’s launch note says the same thing in fewer words — click Sources, open the sidebar, record the references.

If Sources does not appear, you did not watch a live retrieval. You watched a memory answer. Log that as search_off or no_sources. It is a different problem than “they cited a rival.”

Step	What you do	What you write down
1	Paste the frozen prompt. Do not “improve” it mid-run	prompt_id + verbatim text
2	Confirm search / browsing actually ran	search_on or no_sources
3	Open Sources. Copy every URL	Domain + path, not a vibe
4	Read the answer for brand names	Named / not named / wrong fact
5	Screenshot or export the panel	Date + product version if shown

Eligibility is not a mystery either. OpenAI’s crawler docs say OAI-SearchBot is the bot that surfaces sites in ChatGPT search. Sites opted out of that bot are not shown in search answers, though they can still appear as navigational links. The publishers FAQ repeats the instruction: do not block OAI-SearchBot if you want to be in summaries and snippets. A 24-hour lag after a robots.txt change is the documented wait, not a folklore SLA.

• Prompt pasted verbatim from the panel sheet
• Sources panel opened, not guessed from memory
• Every cited URL stored, including directories you dislike
• Your status set to one enum: cited_link, named_only, absent, misrepresented
• robots.txt checked for OAI-SearchBot before you blame the copy

How do I log Perplexity and AI Overviews without an API?

Stay in the consumer products. Do not stand up a paid Perplexity API job to “automate citations.” The brief is who the buyer sees, not who an endpoint returned on a different model path.

Perplexity’s consumer UI shows numbered footnotes in the answer and a Sources list of page titles and domains. Click the numbers. Copy the URLs. Treat “source card at the top” and “footnote on a sentence” as two columns if they diverge — they sometimes do.

Google is a different surface and a different report. AI features documentation says a page must be indexed and snippet-eligible to appear as a supporting link in AI Overviews or AI Mode. There is no extra AI-only markup. Google also describes query fan-out: one prompt can spawn related searches, which is why a Reddit thread or a docs page shows up next to a polished guide.

Search Console will not replace the Overview screenshot. The June 3, 2026 generative AI performance reports give impressions, pages, countries, devices, and dates for AI Overviews and AI Mode. The Search help page is explicit: an impression is a link to your site shown in a generative AI feature. It does not list queries. It does not list ChatGPT. It does not list Perplexity. Rollout is still a subset of properties.

Product	What you can see by hand	What you cannot treat it as
ChatGPT Search	Sources sidebar + inline chips	A Search Console query report
Perplexity	Footnotes + Sources cards	An official citation Search Console
AI Overviews / AI Mode	Visible supporting links + GSC gen-AI impressions	ChatGPT coverage
Semrush / similar	Mention, citation, and source tables on their prompt set	Your frozen panel unless you verify a sample

Build a matrix: prompts on rows, products on columns, your status in the cell. Executives read matrices. They do not read paragraphs about “AI.”

How do I build a competitive prompt panel?

Start from money, not ego. Sales language beats keyword-planner language. Cap v1 at 25–40 prompts so you can re-run it. Expand after the ritual works.

Prompt class	Example shape	Why it belongs
Category recommend	“Best [category] for [ICP]”	This is where directories win
ICP / size / industry	“Best [category] for a 12-person [industry] team”	This is where you should win
Head-to-head	“[You] vs [Rival]” and the two rival-only versions	Comparison pages get cited
Problem / how-to	The question sales hears weekly	Thin blogs are beatable here
Brand	“Is [You] worth it?” / “What does [You] cost?”	Accuracy incidents live here
Ecosystem	“Best [category] Shopify app” if pipeline starts there	Marketplace copy is the page

Freeze 3–6 competitors for 90 days. Include one aspirational owner of citations, one peer, and one cheap substitute. If sales adds a new rival mid-quarter, park them on a watch list. Moving the set mid-stream hides whether the work worked.

• Every prompt maps to a pipeline conversation, not a vanity query
• Competitor set named and dated
• Prompt IDs never change; wording changes get a new ID
• Out-of-ICP prompts marked wontfix up front
• Owner of the sheet is a person, not “marketing”

What columns belong in the citation log?

Use boring columns. Fancy dashboards die in the handoff. A sheet survives.

date | product | prompt_id | prompt_text | search_or_sources | brands_named | urls_cited | our_status | accuracy | ticket | owner

Field	Enum or rule	Why it exists
search_or_sources	on / off / unknown	Memory answers are not citation losses
brands_named	Comma list, including you	Mention ≠ citation
urls_cited	Full URLs, not domains only	The brief is a page, not a brand
our_status	cited_link / named_only / absent / misrepresented	Four tickets, not one mood
accuracy	ok / minor / material	Wrong price is not a win
ticket	page / pr / listing / entity / wontfix	The gap is the backlog

Export a monthly pivot: prompts where our_status=absent and a competitor URL appears at least twice. That pivot is the content and PR brief. Share of voice, if you compute one, is your cited_link count ÷ panel runs against a frozen denominator. Write the formula on the sheet. If you cannot write the formula, you do not have a percentage.

Why does a single run lie?

Answer products are not rank trackers. The same prompt, next Tuesday, can cite a different trio. Semrush says as much in the gap guide: one response tells you very little; you need a pattern. That is why the panel is frozen and the cadence is boring.

What changed	How it shows up	What you do
Search did not run	No Sources control	Log search_off. Do not score a citation loss
Fan-out took a different path	New supporting link, same gist	Count the URL; do not panic on one swap
Personalization / location	Local pack or geo-specific docs	Note locale. Re-run logged-out if the buyer is
Product UI moved	Chips vs sidebar vs footnotes	Still copy the URL. The chrome is not the metric
You “improved” the prompt	A different question got answered	Void the row. Use the ID or make a new one

Procedure before you call a gap real:

1. Run the prompt three times across at least two days, same product, same wording.
2. Union the cited URLs. The brief is the set, not the first screenshot.
3. Mark a domain “owns this prompt” only if it appears in at least two of the three runs or it is the only cite on a money prompt you will fund anyway.
4. Keep the screenshots. Arguments die when the panel is gone.

A competitor that appears once in thirty days is noise. A directory that appears on eleven of twelve category prompts is the brief. Do not average them into one percentage.

What does a mention-only cell actually mean?

named_only is not a consolation prize and it is not a citation. The model used your name and credited someone else’s URL. That is a source gap wearing your brand.

Cell	What happened	Ticket
cited_link	Your URL is in Sources / footnotes	Protect the page. Watch accuracy
named_only	Brand in the prose, no your-domain URL	PR to the cited host, or a page worth crediting
absent	Neither name nor URL	Page and/or PR, scored by winnability
misrepresented	Name or URL with a wrong fact	Fact packet first. Visibility second

Worked row — format only, fill from your sheet:

2026-08-16 | chatgpt_search | p12 | “best [category] for [ICP]” | on | RivalA, RivalB | [roundup host]/best-…, [rival]/compare | named_only | minor | pr | [owner]

That row already is the brief: pitch the roundup, or ship a comparison page that is easier to credit than RivalA’s. Do not write “we have 12% SOV” on top of it.

• Mention column and citation column never get summed into one cell
• named_only always captures the URL that was credited
• Wrong-price name-drops escalate to misrepresented
• Partner-page citations get a decide-to-own note, not a victory lap

How do I turn the log into a content and PR brief?

Read the URL leaderboard, not the brand leaderboard first. Brands get named via partner pages you do not control. URLs tell you what to write and who to pitch.

Leaderboard	How you build it	What you do with it
URL	Count cited URLs across the panel	Out-write or earn the page
Domain	Roll URLs up to hosts	PR list for digital PR
Brand	Count names in answer text	Entity and fact-packet work
Format	List / table / docs / thread	Steal structure, not sentences

Ahrefs’ 75,000-brand analysis reported a 0.664 Spearman correlation between branded web mentions and AI Overview brand visibility, against 0.218 for referring domains. That is a correlation, not a law, and it is Overview-side. It still explains why a sentence on a domain the panel already cites beats another homepage backlink with no quote attached.

For each of the top 10 external URLs:

1. What format got lifted — definition, table, steps, list, docs, thread?
2. How old is the visible date?
3. Are the category facts accurate?
4. Is there a path to inclusion — contribute, earn the roundup slot, or outpublish?
5. Would you be proud to be quoted next to them?

If the winning URL is a thin listicle, outpublishing is viable. If it is a primary standards document, change the prompt strategy. Do not cosplay as ISO.

How do I read a cited URL like a brief?

Open the URL. Find the passage that could have been lifted. If you cannot point at 40–80 words that answer the prompt, you are guessing why it won.

Check	Pass	Fail
Lead sentence	Answers the prompt in one line	Brand story, then the point on paragraph six
Structure	Table, steps, or ICP rows	Unbroken essay
Date	Visible updated or collection window	Undated “ultimate guide”
Entity	Legal name, product name, who-it’s-for	Pronouns and “we” only
Proof	A number, a method, or a named constraint	Adjectives
Inclusion path	Editor, contribute, or outpublish	No public contact and no equivalent page you can ship

Write three lines into the ticket:

1. The lift-ready passage (quote it).
2. The gap — missing ICP, stale date, no criteria, or you are not on the list.
3. The play — page with a named H2, or pr with a named editor and a locked two-sentence fact.

That is the brief. A card that says “improve AEO” is not.

• Passage quoted in the ticket, not paraphrased from memory
• Date on the cited page recorded
• Inclusion path chosen before a writer is assigned
• Standards docs marked unwinnable instead of “write a better ISO”

How do I classify gap types?

Classify before you assign a writer. The wrong ticket wastes a sprint.

Gap type	Signal in the log	Primary fix
Absence	Never cited on category prompts	Answer-first page + earned mention
Substitution	Competitor how-to or docs cited	Clearer method, dates, and proof
Directory dominance	List / roundup hosts win	Pitch those exact URLs; ship your comparison
Stale corroboration	Old press still quoted	Update the source or outpublish it
Identity error	Wrong facts about you	Fact packet + hallucination repair
Crawl block	Sources never include you; robots blocks SearchBot	Fix robots / WAF, then retest
Ecosystem	Answers cite a marketplace listing	Fix the listing, not the blog

Google’s AI-optimization guide still starts with Search technical requirements and people-first pages. Their helpful-content questions still ask whether the page offers original information. A gap analysis that produces “write 40 more posts” without a cited-URL target is SEO cosplay.

• Each absent money prompt has one ticket type
• Source-gap URLs copied into the PR list, not a generic media list
• Identity errors get a fact-packet diff, not a new thought-leadership essay
• Crawl blocks get an engineering ticket before a copy ticket
• wontfix has a written reason

How do I prioritize which gaps to close first?

Score each absent money prompt 1–5 on pipeline influence, buyer frequency, winnability, and strategic fit. Multiply influence × winnability. High influence / low winnability becomes a long PR play, not a weekend blog post.

Score	Influence	Winnability	What you fund
High × high	Sales hears it weekly	Cited URLs are thin or stale	Page this sprint
High × low	Sales hears it weekly	Cited hosts are standards or giant directories	60–90 day PR
Low × high	Rare in pipeline	Easy page to ship	Only if it supports a cluster
Low × low	Vanity	Entrenched	wontfix

Winnability is a reading of the cited URL, not a feeling. A 2019 listicle with no criteria table is winnable. A government spec is not.

Do not prioritize “prompts where a tool showed a low SOV.” Prioritize prompts you can point at in a sales call. The sheet exists to stop leadership from funding the prompt that sounded smart in a Slack thread.

What plays actually close a citation gap?

Five plays. Pick from the ticket type, not from a content calendar theme.

Own the missing answer page. If every citation is a “What is X?” post and you lack one, write it. Lead with the sentence a model can lift. Put a table or steps under the first heading. Date it.

Beat the cited URL on compressability. Tables, numbered procedures, explicit ICP, visible dates. FAQPage markup only if the questions are on the page. Do not match word count. Match extractability.

Earn a seat on the roundup. Digital PR aimed at the exact domains already in the log is the job. Random guest posts on domains the panel never cites are optional SEO. See digital PR for citations.

Fix the listing. App stores, marketplaces, and partner directories get mirrored. Classic web PR will not repair a thin Shopify listing that Perplexity already footnotes.

Repair the fact. A misrepresented cell is not a visibility win. Lock the one- and two-sentence facts, update the page, and pitch the source that taught the model the wrong year.

Ticket	Ship this	Retest after
page	Answer-first URL with a lift-ready lead	2–6 weeks
pr	Placement on a logged domain	4–8 weeks
listing	Marketplace / partner copy aligned to the fact packet	2–4 weeks
entity	SameAs / Organization / visible legal name cleanup	Next full panel
wontfix	Written reason in the sheet	Do not retest for vanity

Steal structure, not sentences. Reverse-engineer H2s, table shape, and FAQ list. Then beat them on ICP boundaries, freshness, original proof, and accurate entity markup. Derivative sameness is visible to retrieval systems, humans, and lawyers.

Why do directories and roundups keep winning?

They compress category advice into a list a model can quote. Google’s AI-features note that Overviews exist to get people to the gist and then out to supporting links. A “best X for Y” page with a criteria table is purpose-built for that job. Your 2,400-word brand essay is not.

Cited format	Why it wins the footnote	Your counter
Best-of list	ICP rows, prices, “who it’s for”	Ship your own comparison with named criteria
Docs / changelog	Dated, specific, hard to reassign	Publish the method, not the vibe
Trade roundup	Third-party voice + many brands	Pitch that editor with a locked fact
Community thread	Specific failure, specific fix	Contribute where the panel already cites
Homepage	Entity fallback when nothing else exists	Do not celebrate this; it is a gap

You will not dethrone a directory by complaining about directories. You either get on the list the panel already cites, or you publish a comparison that is easier to lift than theirs. Both is allowed. Neither is a LinkedIn rant.

When is absence the correct answer?

Sometimes the model omits you because you are out of scope. “Best enterprise suite” when you serve mid-market is not an insult. It is a filter. Forcing inclusion with spam creates bad-fit demos.

Signal	Correct move	Wrong move
Prompt is outside ICP	wontfix + narrow the panel	Doorway page for enterprise theater
You are named via a partner	Decide if you want that URL	Pretend it is “your” citation
Facts are wrong but you are cited	Accuracy ticket first	Celebrate the link
You blocked OAI-SearchBot	Allow the search bot, wait ~24 hours, retest	Rewrite the hero copy
GSC gen-AI impressions exist, chat panel is empty	Two surfaces, two briefs	Average them into one SOV

False inclusion is not a win. A prompt that sends the wrong buyer into a sales call is a cost. Write the reason in the sheet so the argument does not restart every quarter.

What breaks if you skip the log?

The failure mode is theater. Someone pastes one ChatGPT answer in Slack. Leadership funds a “thought leadership” series. Six weeks later the same directory still owns the money prompt, and nobody can prove the series did anything because the prompt wording changed.

Skip	What breaks	What it costs
No frozen panel	Every meeting argues a different question	A quarter of unmeasured pages
No URL column	You pitch “the industry” instead of a host	PR that never gets retrieved
Mention = citation	You celebrate a name-drop with a wrong price	Hallucinated brand facts in the wild
Moving competitors	Charts improve by definition	You cannot tell if the page worked
Treating GSC as ChatGPT	You optimize Overviews and miss the chat panel	Split-brain reporting
Invented SOV	A percentage with no denominator	Budget against a fiction

I have watched this pattern across 500+ automations and hundreds of sites: the teams that keep a boring sheet outrun the teams that bought a dashboard logo. The sheet is the product. The dashboard is optional corroboration.

Store decisions in the sheet, not in chat. Six weeks later nobody remembers why a prompt is wontfix, and the argument restarts.

How should leadership read a gap report?

Lead with money prompts. Show one side-by-side: the cited competitor URL versus the page you do not have. Ask for budget against the gap list, not against “AI” as a vibe. Revisit in 60 days with the same prompts.

Write the slide like this — replace the counts with your sheet, never with a number you guessed:

On [N] recommendation prompts for our ICP, we were cited_link on [A] and absent on [B]. Competitor [Name] appeared via [URL 1] and [URL 2]. We lack [page type]. Proposed: ship [page] in [weeks]; pitch [editor / domain] with the locked fact packet. Retest on [date] with the same prompt IDs.

That paragraph is fundable. A generic AI strategy deck is not.

Leadership ask	Show this artifact	Do not show
Where are we losing?	Absent-money-prompt pivot	A vibe word cloud
What do we fund?	Top 5 tickets with owners	A 40-item editorial calendar
Did it work?	Same prompt IDs, two dates	A new competitor set
What did it cost?	Hours + placements, not “SOV points”	A screenshot of a vendor gauge

Share a one-page dashboard: citation rate on the frozen panel, top gap URLs, shipped tickets, next retest date. If you report share of voice, print the formula and the competitor freeze date on the same slide. If you cannot, cut the percentage.

What cadence keeps the panel honest?

Weekly, monthly, quarterly. Miss a week and the ritual becomes archaeology.

Cadence	What you run	What you ship from it
Weekly	10–15 money prompts	New absent cells → tickets
Monthly	Full panel + URL leaderboard	PR list + page queue
Quarterly	Retire dead prompts; add sales language	Competitor freeze review

Run a 25-minute standup:

1. Three minutes: citation-rate sparkline on the frozen panel
2. Seven minutes: new absent money prompts
3. Seven minutes: assign or advance tickets
4. Five minutes: accuracy incidents
5. Three minutes: next retest commitments

Invite content, PR, and someone from sales at least twice a month so prompt language stays real. Keep owners named. When someone goes on leave, transfer the ritual in the sheet — AEO dies in the handoff.

Week-one kit, four artifacts, or you are still in opinion mode:

• Prompt panel sheet with IDs
• Competitor freeze list with a date
• URL leaderboard pivot
• Backlog tagged page / pr / listing / entity / wontfix

Repeat the kit after major launches. The cost of re-baselining is tiny next to a quarter of unmeasured content. If you want a second pair of eyes, the visibility lane exists for that reason: /visibility and the visibility audit path turn the kit into a managed baseline. Either way, ship the ritual before you buy another dashboard.

FAQ

How do I see what ChatGPT cites?

Open ChatGPT Search, paste a frozen buyer prompt, and record the Sources panel plus any inline citation chips. OpenAI documents hover-and-click citations and a Sources sidebar under the response. Repeat across sessions. Pair the run with Perplexity footnotes and an Overview check. No single export replaces that panel today.

What is an AI citation gap analysis?

It is a structured comparison of the sources and brands answer products cite for your priority prompts versus your own presence — then a prioritized plan to earn inclusion. Mentions, citations, and third-party source gaps stay in separate columns. The output is a content and PR brief, not a lonely percentage.

How many prompts do I need?

Enough to cover revenue-relevant intents without drowning the team. Most brands can learn from 25–40 well-chosen prompts re-run on a weekly subset and a monthly full pass. Add prompts when sales language changes. Do not add prompts to make a chart look busier.

Why do directories keep winning citations?

They compress category advice into lists and tables models can quote. Either earn placement on the quality lists your panel already cites, publish a better-owned comparison with named criteria, or both. Complaining about directories is not a play.

Can Semrush show ChatGPT citations?

Semrush can show mention, citation, and source tables on the prompt sets it tracks, and it publishes those definitions in its AI visibility docs. Treat that as corroboration. Your frozen panel is still the brief unless you have verified that the tool’s prompts match the questions buyers ask you. Search Console gen-AI reports cover Google surfaces, not ChatGPT.

How fast can gaps close?

On-site pages can enter browsing-mode answers within days to weeks after they are crawlable and snippet-eligible. Displacing an entrenched roundup is a PR cycle measured in months. OpenAI documents about a day for robots.txt search-bot changes to take effect. Measure monthly on the same prompt IDs, not daily on new wording.

CTA

Stop guessing who AI quotes. Log the panel, classify the gap, ship the page or the pitch, re-measure.

Lane overview: /visibility. Next step: a visibility audit.