Source gap analysis: the pages AI cites instead of you

When a buyer asks an AI assistant which accounting tool suits a five-person agency, the answer often arrives with a handful of links beneath it: a comparison article on a review site, a competitor's pricing page, a forum thread, a video walkthrough. If none of them is yours, the engine has built its recommendation from other people's evidence. A source gap analysis turns that observation into a repeatable method. It is a structured comparison of the URLs that generative engines cite for a fixed set of prompts, classified by owner and by type, which shows where competitors or third parties are cited and you are not, and what kind of response each gap calls for.

The method has five steps, each producing what the next one needs: a frozen prompt set, a log of cited URLs, a classification of every source, a gap matrix and a ranked list of actions. A short review of published research comes first; the legitimate responses, the illegitimate ones and a test for closure come after.

What exactly counts as a source gap?

For a given prompt, a source gap exists when the answer cites pages that address the buyer's question, none of them is yours, and at least one works in a competitor's favor, because the competitor owns it or because it names the competitor and leaves you out. This is narrower than "we are not in the answer": your brand can appear in the prose while your domain is missing from the sources, as our piece on the difference between being mentioned and being cited explains.

Two variants lead to different work. A counterpart gap occurs when a competitor's own page, such as its pricing page or a blog guide, is cited while your page on the same subject is not. An inclusion gap occurs when a cited third-party page, such as a review-platform category, a directory listing or a roundup article, features competitors but not you. The first is solved mostly on your own site, the second only on someone else's, which changes both the effort and the rules.

What do published studies say about the domains AI engines cite?

Large citation studies diverge in large part because their methods do. A University of Toronto team (Chen, Wang, Chen and Koudas) posted a preprint to arXiv in September 2025 comparing web-enabled AI engines with Google. In one experiment, 1,000 ranking-style consumer prompts were run in the United States and Canada, and a language model classified each cited domain as brand-owned, earned (independent publishers, review and comparison sites) or social. The search-enabled GPT model leaned heavily on earned media: in US consumer electronics, 92.1 percent of the domains it cited were classified as earned and social sources were negligible, while Google kept a more balanced mix. In another experiment, Perplexity surfaced noticeably more social content than the other engines.

Vendor studies add a warning about metrics. Profound, a monitoring vendor, analyzed about 680 million citations from August 2024 to June 2025: Wikipedia, ChatGPT's most-cited source, stood at 7.8 percent of overall citation volume in its tracked data but at 47.9 percent once the denominator was limited to ChatGPT's ten most-cited domains, a contrast that shows how long the tail is. Semrush, tracking more than 230,000 prompts weekly from July to October 2025, saw Reddit fall from close to 60 percent of ChatGPT responses in early August to about 10 percent by mid-September, a shift its authors say the data does not fully explain. Ahrefs reported in March 2026 that 37.9 percent of URLs cited in AI Overviews also appeared within the first ten positions of the results page for the same query, counting ads and SERP features as positions (37.1 percent for organic links alone), against about 76 percent in its July 2025 study. Ahrefs notes that Gemini 3 began powering AI Overviews and that it improved its own parsing method in between, so the decline cannot be attributed to a single cause, but in the new dataset a top-ten ranking was clearly far from a precondition for being cited.

The safe conclusion is modest: third-party pages carry a large share of citations, the mix varies by engine and category, and it moves over time. No study's average tells you which pages are cited for your buyers' questions, so the analysis must start from your own prompts. Why engines seem to trust some sources more is covered in our article on trust and authority signals, and the community side in why AI cites Reddit and forums.

Step 1: which prompts should form the fixed set?

A gap is meaningful only relative to a defined set of questions. A few dozen prompts suffice for a first pass, spread across discovery, comparison and decision stages and written in customers' own words, which our guide to prompt research shows how to collect. Give each prompt a market, a language and a business-importance score, for example from 1 to 5. Then freeze the wording: if it changes between runs, a change in sources may reflect the rephrasing rather than the engine, so track alternative wordings as separate variants.

Step 2: how do you collect and normalize cited URLs?

Engines expose sources differently. Perplexity states that each answer includes numbered footnotes linking to original sources; OpenAI's help documentation says ChatGPT responses using web search may include citations and a Sources view; Google describes supporting links shown with AI Overviews. For every run, log the engine, date, market, language, prompt, citation position and full URL, and repeat each prompt over several weeks, storing failed runs as missing rather than as "no citations". Normalize URLs to registrable domains for counting, but keep the canonical page URL: a review platform may be cited through a category page listing competitors, a competitor's profile or yours, three quite different situations.

Step 3: how should each source be classified?

Classify every cited URL on two independent axes. Ownership asks who controls the page: you, a competitor, an independent third party or users. Type asks what kind of source it is: editorial or news, review platform, directory, marketplace, forum or Q&A, video, social network, encyclopedia or documentation. For third-party pages, also record whether the page features competitors, you, both or neither. Separate axes keep "third party" from becoming one bucket, since an editorial review and a directory listing need different responses; where a type is genuinely unclear, label it undetermined.

Step 4: how do you compute where competitors appear and you do not?

The gap matrix is set arithmetic. For each prompt, collect across its runs the URLs you own, the competitor-owned URLs and the third-party URLs featuring competitors but not you. A non-empty competitor set alongside an empty set of your own marks a counterpart gap; a non-empty third-party set marks an inclusion gap. Pivoting by domain then usually shows a few domains accounting for most inclusion gaps.

Consider a hypothetical: a project-management vendor tracking 40 prompts finds one review platform cited in 14 of them, typically through a category page where three competitors have complete profiles and its own is unclaimed. A competitor's pricing comparison is cited in nine decision-stage prompts, while the vendor's own pricing page never appears. These are two gaps with two owners, and the matrix makes that division of labor visible before work starts.

Step 5: how do you decide which gaps to close first?

Score each gap on three factors and multiply them, so a weak score on any one pulls the total down. Frequency is the share of prompts and runs in which the gap appears; once is an observation, persistence across weekly runs is a pattern. Business importance is the prompt's score, so decision-stage gaps outrank loosely related ones. Feasibility is how much of the outcome you control: your own page fully, a directory profile mostly, an independent roundup hardly at all. The product serves ordering, not precision; confirm gaps in volatile source types over more runs before investing.

What is the legitimate response for each source type?

Engines do not publish selection weights, so the second column is a reasoned interpretation rather than a documented mechanism, and the time horizons are planning estimates.

Source type	Why an engine plausibly cites it	Legitimate action	Effort and horizon
Your counterpart page	A specific answer from the party with first-hand knowledge	Rewrite it to answer the prompt explicitly, with verifiable facts	Low to medium; weeks
Competitor-owned page	It answers a question your site leaves open	Publish or improve your equivalent page; never copy theirs	Medium; weeks
Third-party editorial	Independent comparison with an editorial filter	Earn coverage with real news, data or expertise; disclose sponsorship	High; months
Review platform	Many first-hand opinions in comparable form	Complete the profile; ask every customer for an honest review	Medium; months
Directory or marketplace	Structured facts: category, location, price, availability	Claim and complete listings; keep them consistent with your site	Low to medium; days to weeks
Forum or Q&A community	Experience in the buyer's own words	Participate as yourself, disclose affiliation, follow community rules	High; ongoing
Video platform	Demonstrations and spoken explanations	Publish useful demonstrations with accurate titles and captions	Medium to high; months
Encyclopedia	Neutral, cross-referenced reference text	Earn independent coverage; never edit about yourself undisclosed	Largely outside your control
Documentation	Precise procedures for how-to questions	Publish complete, crawlable docs and changelogs	Medium; weeks

Counterpart gaps are the cheapest to close because the page is yours: answer the prompt better than the cited page, state the answer early and include specifics a model can quote without context, as described in why AI skips product pages. Where the cited page rests on data you lack, original data and research is often the durable answer. Inclusion gaps sit on pages you do not control, so the route is to become worth including: newsworthy material and honest pitching for editorial roundups, as in our guide to ethical link building; complete profiles and even-handed review requests for platforms; demonstrations worth watching for video, examined in our piece on video in AI answers. Encyclopedias, covered in Wikipedia and Wikidata as brand sources, respond mainly to the independent coverage their editors cite.

Which responses cross the line?

Because third-party sources carry weight, the temptation to manufacture them is strong, and the platforms prohibit it explicitly. Google Maps' user-generated content policy bans reviews not based on real experience, paid or incentivized reviews, conflict-of-interest content, and selectively soliciting positive reviews or discouraging negative ones. The US Federal Trade Commission's rule on consumer reviews and testimonials, in force since October 21, 2024, prohibits fake reviews, buying reviews conditioned on sentiment and undisclosed insider reviews; in the EU, the Omnibus Directive (2019/2161) added submitting or commissioning false consumer reviews to the commercial practices banned in all circumstances under the Unfair Commercial Practices Directive, with national rules applying from May 28, 2022.

Communities and reference works have equivalent rules. Reddit's rules ask users to participate authentically, avoid content manipulation and not impersonate an entity deceptively; the Wikimedia Terms of Use, which have required disclosure of paid editing since 2014, now oblige paid contributors to disclose every employer, client, intended beneficiary and affiliation; Google's spam policies treat buying or selling links for ranking purposes as link spam, while paid advertising or sponsorship links are acceptable when qualified with the sponsored or nofollow attribute. Each rule protects the property that makes a source worth citing, so manipulation defeats itself as well as creating risk. Check the current version of each policy, and take legal advice for your market, before planning third-party activity.

A source gap is closed by becoming worth citing, never by imitating the signals that make a source worth citing.

How do you know whether a gap has closed?

Re-run the frozen prompt set on the same schedule, markets, languages and engines. A gap counts as closed when your URL appears across several runs, not once, and gap counts should be read alongside your overall position, as described in our article on share of voice in AI search. Avoid causal language: if your page enters the sources after a rewrite, the rewrite is a reasonable suspect, not a proven cause, because the engine and competing pages changed in the same period.

In Rocketly GEO Suite, Source Intelligence shows which domains engines cite for your tracked prompts and your position among them, split into own, competitor, third-party and user-generated sources and classified by type. A URL-level explorer keeps each source's history, lists competitor cited pages and shows your counterpart page, meaning your pages seen in the same prompts, which is not proof they were cited in the same answer. GEO Opportunities lists source gaps and community gaps among other gap types, each of which can become a CRM task; the source work brief is a template for your team, and the product never publishes or does outreach itself. Google AI Overview is measured weekly as standard, ChatGPT and Perplexity are in a controlled rollout, and Gemini, Claude and Grok are not measured.

To run this analysis weekly instead of rebuilding it by hand each quarter, create a Rocketly account and add GEO Suite: you set the prompts and their business-importance scores, and the cited sources, their history and the resulting tasks stay next to your CRM.

Frequently asked questions

Should I analyze domains or individual URLs?

Both, in that order. Domain counts show which platforms matter across your prompt set, while URLs show whether a platform is cited through a page featuring competitors, featuring you or neutral. Most actionable decisions, such as which profile to complete or which editor to approach, depend on the URL level.

Is it acceptable to ask customers for reviews on platforms AI engines cite?

Yes, if the request is honest and even-handed. Google's policy allows businesses to encourage genuine reviews but prohibits incentives, selective solicitation of positive reviews and discouraging negative ones. Ask every customer after a real interaction, offer nothing in return, and accept the reviews you receive.

How often should the analysis be repeated?

Measure weekly or close to it, and revisit priorities monthly or quarterly. Platform-level citation shares have shifted within weeks in the studies above, so a single snapshot ages quickly, and a steady cadence separates persistent gaps from one-off appearances.