AI Citations and Brand Mentions Diverge: Evidence From 1,787 Answers
In a fixed panel of 12,509 answer-vendor pairs, 15.6% of pairs with a vendor-domain citation did not mention that vendor. Among pairs with a vendor mention, 70.4% did not cite the vendor's own domain. Citation, mention, and recommendation are separate signals, so measuring only one leaves most of the evidence chain unseen.
Methodology
This observational panel uses Foglift's own AI-search category workspace and no customer data. We analyzed every stored answer from 2026-08-14 00:00 UTC through 2026-08-29 00:00 UTC across ChatGPT, Claude, Gemini, Google AI Overview, and Perplexity. Seven fixed category vendors were crossed with 1,787 answers to create 12,509 answer-vendor pairs. A mention means the normalized vendor domain appeared in the stored competitor_mentions array. A citation means at least one normalized citation URL used that vendor's exact first-party domain. Multiple URLs on the same vendor domain within one answer counted once. Foglift was excluded from the entity set. Public results are pooled and engine-level aggregates only; no vendor standings, prompt text, answer text, customer data, or Foglift visibility result is published.
Do citations and brand mentions measure the same thing?
No. Across 1,787 answers and seven fixed vendors, we observed 1,183 answer-vendor pairs with a citation to the vendor’s own domain. Of those, 184 pairs, or 15.6%, did not mention the vendor in the answer. We also observed 3,377 pairs with a vendor mention. Of those, 2,378 pairs, or 70.4%, did not cite that vendor’s own domain.
The two percentages use different denominators. The 15.6% rate asks what happened after a first-party vendor domain was cited. The 70.4% rate asks what happened after a vendor was mentioned. They should not be added, subtracted, or treated as competing shares of one total.
Three signals, three questions
Citation
Which source did the answer surface?
A citation is a source-level event. It can support the category, a comparison, a factual claim, or the named vendor.
Mention
Which vendor entered the answer?
A mention is an entity-selection event. It says the vendor appeared, without proving endorsement or identifying the source that supported it.
Recommendation
Which vendor fits the buyer's request?
Recommendation requires answer context such as position, positive or cautionary language, sentiment, and the buyer's stated constraints.
The pooled result
| Measure | Pairs | Interpretation |
|---|---|---|
| Vendor domain cited | 1,183 | Denominator for cited-without-mention rate |
| Vendor cited and mentioned | 999 | 84.4% of cited pairs |
| Vendor cited without a mention | 184 | 15.6% of cited pairs |
| Vendor mentioned | 3,377 | Denominator for mentioned-without-own-domain-citation rate |
| Vendor mentioned without own-domain citation | 2,378 | 70.4% of mentioned pairs |
The split by engine
The pooled result hides substantial engine variation. Perplexity had 141 cited-without-mention pairs, compared with three on ChatGPT. The opposite mismatch was common on every engine: 50.6% to 77.3% of mentioned pairs lacked a citation to the vendor’s own domain.
These rates are workflow signals, not an engine quality ranking. Citation availability, retrieval behavior, presentation, and extraction differ by product. The frozen end date also matters. It precedes the August 29 change after which the current Gemini monitoring lane no longer exposes citations.
| Engine | Answers | Cited pairs | Mentioned pairs | Both | Cited only | Cited-only rate | Mentioned only | Mentioned-only rate |
|---|---|---|---|---|---|---|---|---|
| ChatGPT | 349 | 361 | 724 | 358 | 3 | 0.8% | 366 | 50.6% |
| Claude | 359 | 146 | 582 | 142 | 4 | 2.7% | 440 | 75.6% |
| Gemini | 359 | 170 | 677 | 154 | 16 | 9.4% | 523 | 77.3% |
| Google AI Overview | 360 | 196 | 681 | 176 | 20 | 10.2% | 505 | 74.2% |
| Perplexity | 360 | 310 | 713 | 169 | 141 | 45.5% | 544 | 76.3% |
What each mismatch tells you to do
| Observed state | What it establishes | Next investigation |
|---|---|---|
| Cited and mentioned | The vendor and its first-party evidence both entered the answer. | Check position, wording, sentiment, and whether the citation supports the vendor claim. |
| Cited without mention | The vendor domain supplied evidence or category context without entering the named shortlist. | Inspect the cited page’s entity clarity and the independent sources that validate buyer fit. |
| Mentioned without own-domain citation | The vendor entered the answer while the evidence layer pointed elsewhere or exposed no matching first-party URL. | Map third-party sources, verify factual accuracy, and protect the independent corroboration that supported selection. |
A citation dashboard alone cannot measure brand selection
A source-only dashboard misses the 2,378 pairs where a vendor was named without a citation to its own domain. A mention-only dashboard misses the source layer that explains whether the answer relied on the vendor, a review site, a publisher, or another source. A recommendation score built from either field alone collapses different events into one number.
Foglift keeps prompt-level mentions, citations, competitors, positions, sentiment, and dated answer history separate across five engines on paid plans. It then connects that answer evidence to Technical Audits, prioritized recommendations, crawler-purpose analytics, and recognized AI referrals. Launch starts at $49 per month, supports monitoring up to daily, and includes REST API, CLI, and MCP access. See the exact contracts on the AI monitoring page, the competitor methodology, and the developer specification.
Limitations
- The panel covers one AI-search software category and seven fixed vendors. It does not estimate a universal rate across industries.
- Mention detection uses Foglift’s normalized entity contract. Unmapped aliases can create false negatives.
- Citation matching requires the vendor’s exact first-party domain. A publisher, review site, community thread, or social page can support a vendor mention without counting as its own-domain citation.
- The data is observational. It does not show that a citation caused a mention or that a mention caused a recommendation.
- Product-level differences in citation display and extraction can influence the engine split. Cross-engine differences should not be interpreted as causal performance rankings.
Reproducibility and source notes
The aggregation freezes the window, workspace, seven-domain entity set, mention predicate, and exact-domain citation predicate. The downloadable CSV contains pooled and engine-level aggregates. It excludes vendor standings, prompts, answer text, full source URLs, customer data, andFoglift’s own visibility result.
Provider documentation supports the measurement boundary: OpenAI separates search discovery from potential training and user-directed retrieval, Anthropic documents separate training, search, and user-fetch crawlers, and Google describes AI Overviews and AI Mode as grounded Search experiences that surface supporting links. Those retrieval contracts do not turn a citation into a recommendation.
Sources and further reading
Frequently Asked Questions
What is the difference between an AI citation and a brand mention?
A citation identifies a source URL used or surfaced with an AI answer. A brand mention records whether the answer names a vendor. The two can occur together, separately, or neither can occur. This study matches citations only to the vendor's exact first-party domain, so a third-party review can support a mention without counting as a vendor-domain citation.
Does a citation mean an AI engine recommends the cited brand?
No. A citation shows evidence use or source surfacing. The answer may cite a vendor page while selecting another vendor, use the page for category context, or cite it in a cautionary comparison. Recommendation strength requires answer-level context such as position, wording, sentiment, and buyer fit.
Why did 70.4% of mentioned pairs lack a citation to the vendor's own domain?
The result shows that brand selection and first-party evidence are frequently decoupled in this panel. A model may rely on review sites, publisher comparisons, community discussions, or other sources when naming a vendor. The aggregate cannot identify which outside source caused a mention, and the study makes no causal claim.
Can the engine-level rates be treated as a ranking of AI engines?
No. Engines differ in how they retrieve, present, and expose citations. The engine table is a measurement boundary and a workflow guide. It is not a quality ranking, and rate differences should not be interpreted as proof that one engine reasons better or retrieves more accurately.
Why does the study stop before August 29, 2026?
Foglift's Gemini monitoring lane stopped exposing grounded citations after its August 29 instrumentation change. Freezing the panel before that boundary keeps the citation field comparable within the study. A later replication would need a separate design rather than silently mixing incompatible citation contracts.
What should an AI visibility platform measure together?
At minimum, preserve the full answer, brand mentions, positions, citations, competitors, sentiment, engine, prompt, and timestamp. Crawler requests and recognized AI referrals belong in separate evidence layers. A crawler fetch shows access, while a referral shows that a person clicked through from an AI answer.
Free tool
Run a Technical Audit on your own site
Check technical SEO, performance, security, accessibility, and structural AI Readiness. No signup required.
Run Free Technical Audit