AI Citation Gap Analysis: Find the Sources Competitors Win

When a competitor appears in an AI answer and your brand does not, the visible brand difference is only the start of the diagnosis.

The answer may cite the competitor's product page. It may cite an independent review that never included you. It may use your research while naming the competitor. It may cite no source at all. Or the entire difference may disappear on the next matched Run.

Short answer: Run citation gap analysis at the level of Question × answer-engine surface × Run. Freeze the brand and competitor identities, governed domains, market, language, eligible denominator, and coding policy. Normalize cited URLs, classify source ownership and answer role, and separate owned-source gaps, earned-source gaps, factual conflicts, retrieval unknowns, and ordinary volatility. Prioritize repeated high-intent gaps whose evidence and ownership support a legitimate action; do not turn every competitor citation into a page-writing task.

This guide turns raw citation differences into a source backlog that content, SEO, product-data, communications, and partnership teams can review.

A Citation Gap Is A Defined Evidence Difference

Avoid using "citation gap" as a loose synonym for poor AI visibility.

A useful gap statement looks like this:

Across three matched US-English Runs of 12 unbranded enterprise-security Questions on the named ChatGPT search surface, seven completed eligible answers cited two competitor-owned security pages while none cited our governed security or trust URLs. Five of the seven answers recommended those competitors. The target pages were public, canonical, and current at the time of observation.

That statement exposes:

• Question class.
• Surface.
• Market and language.
• Number of matched Runs.
• Completion denominator.
• Competitor and target source rules.
• Answer role.
• Technical eligibility check.

"Competitors get cited more" exposes none of those things.

Citation, Mention, Recommendation, And Click Are Different

The cited-but-not-mentioned visibility funnel separates crawl, retrieval, citation, answer absorption, brand mention or recommendation, and click.

Use the same separation here.

A citation gap can exist without a mention gap. A mention gap can exist even when your page is cited. Diagnose the first observable stage that differs from the intended outcome.

Start With The Experimental Unit

Use one raw unit:

Question × answer-engine surface × Run

Each scheduled unit should end in one governed outcome:

• Completed and eligible.
• Completed but metric-ineligible.
• Failed.
• Refused.
• Blocked or otherwise unavailable.

Do not silently code failed answers as brand absence. Do not add every retry as a new observation. Keep retry evidence and retain the result selected by the predeclared policy.

For each unit, preserve:

• Exact Question and intent class.
• Provider and product surface.
• Market, language, device, and relevant account state.
• Requested and returned model when exposed.
• Search or retrieval mode when exposed.
• Attempt status and eligibility.
• Full answer text.
• Displayed citation URLs, titles, and positions when available.
• Brand, competitor, recommendation, and source coding.
• Observation time, reviewer, and coding version.

The AI search query set guide explains how to govern branded, unbranded, problem, comparison, proof, and buying-stage Questions.

Freeze Identity Before Counting Sources

Citation analysis fails when brand and domain rules move during review.

Create a governed identity table:

Decide how to handle:

• Parent and subsidiary domains.
• Documentation subdomains.
• Regional sites.
• Acquired products.
• App marketplaces.
• Authorized resellers.
• Community and support forums.
• Tracking and redirect URLs.
• Syndicated copies.

Do not classify a third-party article as competitor-owned merely because it praises the competitor. Ownership and editorial stance are different fields.

Normalize Every Cited URL

The same source can appear through multiple URLs.

Store both the displayed URL and a normalized analysis key.

Normalization may include:

• Lowercasing the host.
• Removing a fragment.
• Removing known tracking parameters.
• Resolving an observed redirect to its final URL.
• Recording the declared canonical.
• Mapping mobile or print variants when verified.
• Preserving locale and material page differences.

Do not collapse two pages merely because their titles look similar. Do not assume a canonical chosen by one search engine controls every answer engine. Keep the transformation log so another reviewer can reconstruct the raw source.

Useful fields:

displayed_url, final_url, declared_canonical, normalized_url_key, registrable_domain, source_owner_id, normalization_rule, checked_at

Classify The Source Layer

Use source types that map to real owners and actions.

Source type is not a quality score. A community post can be accurate; an owned page can be outdated. Add separate fields for accuracy, freshness, relevance, evidence strength, and conflicts of interest.

The Seven Citation-Gap Types

1. Owned-Source Absence

The answer cites competitor-owned or third-party pages while no governed target-owned page appears.

Possible explanations include:

• No target page answers the Question.
• The target page is inaccessible, non-canonical, or weakly linked.
• The source is available but not selected.
• The answer surface uses a different source mix on this Run.

Do not label the cause until you inspect the source and technical evidence.

2. Source Exists But The Brand Is Not Named

A target-owned page appears as a citation, yet the answer does not identify the brand.

This can happen when a page supplies a definition, dataset, or process that is absorbed into generic answer language. Review entity attribution, authorship, page context, and the claim-to-source relationship without assuming a markup change will cause a mention.

3. Competitor-Only Recommendation Support

The answer recommends one or more competitors and cites sources that support those recommendations, while the target is absent or unsupported.

Separate:

• Competitor-owned evidence.
• Independent comparative evidence.
• User-generated experiences.
• Public product or pricing data.

The legitimate response depends on which layer shaped the answer.

4. Earned-Source Inclusion Gap

An independent comparison, directory, analyst page, or partner page is repeatedly cited and includes competitors but not the target.

Ask:

• Is the target genuinely eligible for inclusion?
• Is the article current?
• Are its comparison facts accurate?
• Does the publisher accept corrections, updates, or submissions?
• Would outreach create a conflict or violate editorial policy?

The action may be an evidence-backed update request. It is not permission to buy undisclosed placement or manipulate a community.

5. Factual Conflict Gap

The answer or cited source contains inaccurate target information, such as old pricing, unsupported availability, a missing integration, or the wrong company identity.

Use the wrong AI answer correction guide to trace the claim, fix authoritative sources, request appropriate discovery, and validate the correction through matched Runs.

Urgent legal, safety, security, or customer-harm issues should follow the organization's incident and correction process rather than waiting for a monthly citation trend.

6. Citation Concentration Gap

One domain or source type supports a large share of the target's observed citations.

Concentration can make visibility fragile if that page changes, becomes inaccurate, or disappears. Report the concentration; do not assume artificial diversification is automatically better.

A simple descriptive metric is:

top-domain citation concentration = target citations from most-cited domain ÷ all observed target citations

State whether repeated citations to the same URL inside one answer count once or multiple times. Binary per answer is usually easier to interpret for visibility coverage.

7. Volatility Or Coverage Gap

A competitor-only citation appears once and disappears in the next matched observation.

That may be normal answer variance rather than a durable source advantage. The AI search volatility guide explains why completed denominators, repeated Runs, and observed ranges matter before action.

Metrics That Keep The Denominators Visible

Do not collapse the audit into one citation-gap score.

Target-Owned Citation Coverage

target-owned citation coverage = eligible answers with at least one governed target-owned citation ÷ completed eligible answers

This answers:

In how many observed answers did a target-owned source appear?

Competitor-Owned Citation Coverage

For fixed competitor B:

competitor-owned citation coverage(B) = eligible answers with at least one governed B-owned citation ÷ completed eligible answers

Keep each competitor separate before summarizing the cohort.

Competitor-Only Owned-Source Gap Rate

competitor-only owned-source gap rate = eligible answers with competitor-owned citation and no target-owned citation ÷ completed eligible answers

This is a source-presence metric. It does not say the target deserved a citation or that the competitor's page caused the recommendation.

Cited-But-Not-Mentioned Rate

cited-but-not-mentioned rate = answers with target-owned citation and no target brand mention ÷ answers with target-owned citation

The denominator is answers with a target-owned citation, not all attempted Questions.

Repeated Gap Coverage

For a predeclared minimum of matched Runs:

repeated gap coverage = Questions meeting the repeated-gap rule ÷ eligible Questions with enough matched Runs

Publish the rule. For example, "competitor-only owned-source gap in at least two of three matched Runs" is an analyst threshold, not an industry standard.

Keep raw counts beside every percentage.

A Step-By-Step Citation Gap Workflow

Step 1: Declare The Measurement Contract

Freeze:

• Business objective.
• Question set and intent classes.
• Provider surfaces.
• Markets and languages.
• Device and account conditions where material.
• Target identity, aliases, and governed domains.
• Competitor cohort and domains.
• Attempt, retry, and eligibility rules.
• Citation and recommendation coding.
• URL normalization policy.
• Review period and time zone.

Use a version number. If the product surface or cohort changes materially, start a new baseline or record a contract break.

Step 2: Collect Complete Answer Evidence

For every scheduled unit, preserve:

• Status.
• Full answer.
• Source list.
• Source order when the surface exposes it.
• Brand and competitor mentions.
• Recommendation or comparison role.
• Screenshot or raw artifact according to policy.

Do not collect only the "bad" answers. Selective evidence makes the denominator unknowable.

Step 3: Normalize And Classify Sources

Resolve URLs carefully, map ownership, and classify source type. Add review states:

• Verified.
• Probable.
• Unknown.
• Conflicted.

Never automate an ownership conclusion from domain text alone when the entity relationship is ambiguous.

Step 4: Review Claim-To-Source Roles

A displayed citation does not prove every nearby claim came from that page.

For high-priority answers, record:

• Claim or answer passage.
• Displayed source.
• Whether the source contains relevant support.
• Whether the source contradicts the answer.
• Whether the brand is named or merely supplies generic evidence.
• Reviewer confidence.

The 2026 preprint From Citation Selection to Citation Absorption offers a useful distinction between being selected as a citation and appearing to influence answer content. Its measurements are observational proxies, not direct access to hidden model attention. Use the distinction as a review framework, not as proof of causality.

Step 5: Calculate Native Gap Metrics

Calculate coverage separately by:

• Question intent.
• Provider surface.
• Market and language.
• Competitor.
• Source type.
• Owned versus third-party evidence.
• Run and time window.

Avoid averaging channels before inspecting them. A healthy Perplexity citation pattern can hide a repeated ChatGPT source gap.

Step 6: Inspect The Winning Source

For each repeated high-priority gap, review:

• What exact buyer need the source answers.
• Which facts, examples, comparisons, or product attributes it contains.
• Whether those facts are accurate and current.
• Whether it is original reporting, an aggregation, or a copied source.
• Who owns and can change it.
• Which target source should legitimately address the same need.
• Whether the target is actually a relevant answer.

Do not reverse-engineer the cited page into a superficial clone. Build the source the buyer and evidence gap require.

Step 7: Assign One Evidence-Matched Action

"No action" is a valid outcome when the evidence does not support intervention.

Step 8: Validate And Rerun

After an approved change:

• Confirm the final URL returns the intended public content.
• Check canonical, internal links, sitemap, and relevant crawler access.
• Verify the corrected fact or source relationship.
• Record the deployment and discovery-request time.
• Wait the predeclared interval.
• Repeat the matched Questions and surfaces.
• Compare completion, citations, mentions, recommendations, and source roles.

An improved result is an association unless the design supports a causal claim. Continue monitoring across normal model and answer variation.

Use Provider Reports Without Treating Them As Prompt Logs

Microsoft's Bing AI Performance announcement describes total citations, average cited pages, sampled grounding queries, page-level citation activity, and trends across supported AI experiences.

Those signals can identify pages and retrieval themes worth investigating. Microsoft explicitly warns that citation counts do not establish placement, authority, ranking, or a page's role inside an individual answer.

Google's Search generative AI performance report exposes owned-page impressions by page, country, device, and date for supported features. It does not expose the dedicated query and full-answer evidence needed for a competitor citation-gap audit.

Use provider reports to find candidate pages and themes. Use matched answer evidence to determine which brand, competitor, and source gap actually occurred.

Prioritize Gaps Without An Opaque Score

Use a decision table.

If a team needs a queue order, label each dimension and discuss the tradeoff. Do not hide judgment inside a score whose weights no one can explain.

A CSV Schema You Can Audit

Use one row per displayed source per completed eligible answer, plus answer-level rows for answers with no citations.

task_id, run_id, question_id, question_text, intent_class, surface, mode, market, language, device, account_state, attempt_status, metric_eligibility, target_mentioned, target_recommended, competitor_ids, competitor_roles, displayed_url, final_url, declared_canonical, normalized_url_key, registrable_domain, source_owner_id, source_type, source_position, claim_excerpt, claim_support_verdict, target_gap_type, review_confidence, observed_at, raw_answer_url, screenshot_url, reviewer, coding_version, action_owner, action_status

For a no-citation answer, leave the source fields null and preserve citation_count = 0. Do not drop the answer from the dataset.

Common Mistakes

Do not call every competitor mention a citation gap.

Do not treat a displayed source as proof that it supports every answer claim.

Do not count failed answers as target absence.

Do not compare competitors without frozen aliases and governed domains.

Do not classify favorable third-party coverage as competitor-owned content.

Do not collapse redirected, canonical, regional, and syndicated URLs without a transformation log.

Do not build one blended citation score across providers with different surfaces and reporting rules.

Do not copy the cited competitor page.

Do not buy undisclosed coverage, fabricate reviews, or manipulate communities.

Do not act on a volatile gap without matched repetition, unless a material factual or safety issue requires immediate correction.

Do not present a post-change citation increase as causal proof.

The Bottom Line

AI citation gap analysis should explain which source layer is missing, which competitor or third party currently supplies the evidence, how often the pattern repeats, and which authorized action follows from that evidence.

Start with stable Questions and complete outcomes. Normalize URLs. Govern brand and competitor identities. Separate citation from mention and recommendation. Classify ownership before assigning work. Then rerun the same measurement contract after a focused change.

The goal is not to force the brand into every answer. It is to make strategically relevant source gaps visible, reviewable, and actionable without confusing correlation with causality.

Create an AEO Table account to preserve Tasks, repeatable Runs, competitor appearances, citations, and source evidence before turning gaps into content, product-data, technical, or earned-source work.