LLM Citation Analysis
Find the exact chunk that gets your competitor cited.
We run your corpus against the models that matter, capture every citation at the URL and passage level, and reverse-engineer the retrieval decision. You get the failing chunks, the fix, and a number you can put in a board deck.
6 models · 400+ prompts · n≥30 samples per prompt · 10 business days
- models sampled — GPT, Gemini, Claude, Perplexity, Copilot, AI Overviews
- 6models sampled — GPT, Gemini, Claude, Perplexity, Copilot, AI Overviews
- margin of error on citation share at n=30 per prompt
- ±4.1%margin of error on citation share at n=30 per prompt
- from kickoff to the passage-level fix list
- 10 daysfrom kickoff to the passage-level fix list
Failure taxonomy
Six reasons a model skips you. We identify which ones are yours.
Every uncited prompt in your run is classified against this taxonomy, with the offending URL and passage attached. No “improve your content” advice.
F-01 · diagnosis
Chunk not self-contained
Your passage answers the question only in the context of the paragraph above it — pronouns, 'this approach', 'as mentioned'. Retrieved in isolation it resolves to nothing, so the model discards it and takes a competitor's standalone paragraph instead.
F-02 · diagnosis
Entity ambiguity
The model can't confidently tell your brand apart from a similarly-named company, or can't resolve you to a single entity across your own site. Ambiguity is a safety risk, so it names someone it is sure about.
F-03 · diagnosis
Corroboration deficit
You state the claim; nobody else does. Retrieval-augmented models weight independent agreement heavily, so a single-source claim on your own domain loses to a claim repeated on three review sites and a trade publication.
F-04 · diagnosis
Crawler exclusion
GPTBot, ClaudeBot, PerplexityBot or Google-Extended are blocked, throttled, or served a JS shell with no content. You are not losing the comparison — you were never in it.
F-05 · diagnosis
Freshness decay
Your page carries no reliable date signal or the content is genuinely stale. On any prompt with a temporal cue — 'latest', 'current', 'this year' — the model prefers a dated competitor page it can defend.
F-06 · diagnosis
Format mismatch
The answer exists on your site but as a video, a PDF, an image of a table, or a JS-rendered accordion. It is not extractable text at the moment of retrieval, so it may as well not exist.
Methodology
A measurement pipeline, not a monthly opinion
- 01
Prompt set construction
We build 400+ prompts from your commercial intent map — comparison, alternatives, pricing, integration, category and objection queries — weighted to revenue rather than volume.
400+ prompts · intent-weighted · client-approved - 02
Multi-model sampling
Each prompt is run n≥30 times per model at controlled temperature and locale, because a single run tells you nothing about a stochastic system.
6 models · n≥30 · locale + persona held constant - 03
Citation extraction
Every response is parsed for brand mentions, cited URLs and citation position. Domains are resolved and normalised so aggregators aren't miscounted as competitors.
URL-level · position-aware · domain normalised - 04
Chunk attribution
For each citation we identify the specific passage the model most likely retrieved, and compute embedding similarity between the prompt, your chunk and the winning competitor chunk.
passage-level · cosine similarity delta - 05
Log-file forensics
Server logs are segmented by agent user-agent to establish what each crawler actually fetched, what it got back, and how often it returned.
GPTBot · ClaudeBot · PerplexityBot · Google-Extended - 06
Remediation modelling
Failures are classified, effort-scored and ranked by projected citation-share gain, so the fix list is ordered by return rather than by ease.
ICE-scored · projected Δ citation share
| Model | Baseline | Day 90 | Movement | Primary driver |
|---|---|---|---|---|
| GPT (browsing) | 12.4% | 47.1% | F-01 chunk rewrites | |
| Gemini | 18.0% | 44.6% | F-05 date signals | |
| Claude | 9.2% | 38.3% | F-03 corroboration | |
| Perplexity | 22.7% | 61.5% | F-04 crawler access | |
| Copilot | 14.1% | 35.9% | F-02 entity graph | |
| AI Overviews | 8.6% | 29.4% | F-06 extractability |
Same corpus, same prompt set, 90 days apart. The only variable was the remediation work.
Benchmark my domainWhat lands in your inbox
Seven artefacts. Each one is machine-readable or your team can ship it.
- CSV + JSON
Prompt-level citation matrix
Every prompt × model × run, with brand mention rate, cited domains and your position in the citation list.
For your analytics team - Redline doc
Passage-level rewrite list
The failing chunks, quoted, with the rewritten version beside them and the failure code that triggered it.
For your content team - Ticket-ready
Technical fix backlog
Rendering, crawler access, sitemap and date-signal issues written as tickets with acceptance criteria.
For your developers - JSON-LD
Entity and schema package
Organization, Product, FAQ and sameAs markup, validated, ready to paste into your CMS or templates.
For your CMS owner - Ranked plan
Corroboration roadmap
The specific third-party properties that would most change your citation odds, ordered by effort against impact.
For your PR lead - Dashboard
Citation share tracker
Your baseline and monthly re-runs on the same prompt set, with competitor movement flagged.
For your CMO - 60 minutes
Engineering readout
A working session with the analyst who ran it — not an account manager reading a summary.
For everyone above
Proof
What the numbers did after remediation
- Enterprise SaaS
Lumen
Cited in 12% of category prompts despite ranking top-three organically. Chunk attribution showed comparison content buried under a JS accordion — invisible at retrieval. Rendering fix plus 22 rewrites.
12% → 47%Citation share90 days - Fintech
Northbay
Two entities with near-identical names in the same market. We consolidated the entity graph, aligned 14 third-party records and rebuilt the sameAs set. Ambiguity failures went to zero.
0F-02 failures remaining120 days - Industrial
Atlas+Co
GPTBot and PerplexityBot were being 403'd by an over-tuned WAF rule. Nobody had checked. Log forensics found it in week one; citations appeared within a single crawl cycle.
+318%Cited prompts60 days
Investment
Measure once, or engineer the whole retrieval surface
Fixed-scope audit, or an ongoing programme with the remediation work included. Month-to-month after the first quarter.
- Fixed scope
Citation Audit
One rigorous baseline you can act on, with no ongoing commitment.
$14,500one-off · 10 business days- 400+ prompts across 6 models
- n≥30 sampling per prompt
- Chunk-level failure attribution
- Crawler log forensics
- Full deliverable set (7 artefacts)
- 60-minute engineering readout
- OngoingMost chosen
Citation Programme
The audit plus the team that ships the remediation and re-measures it monthly.
from $6,500per month · 90-day term- Everything in the Citation Audit
- Monthly re-run on the same prompt set
- Content and schema work delivered
- Corroboration and digital PR outreach
- Competitor citation displacement
- Monthly analyst review call
- Enterprise
Retrieval Engineering
Multi-brand or multi-market estates where retrieval is a platform problem.
Customscoped to estate- Everything in the Citation Programme
- Multi-market and multilingual prompt sets
- Log pipeline into your own warehouse
- API access to the citation dataset
- Embedded analyst and content lead
- Quarterly board-level reporting
Before you ask
The technical questions that decide it
Answer
How do you handle model non-determinism?
By sampling, not by asserting. Every prompt is run at least thirty times per model with temperature, locale and persona held constant, and we report citation share as a rate with a confidence interval — ±4.1% at n=30. Anyone showing you a single screenshot of one chat session is showing you noise. Re-runs use the identical prompt set so month-over-month deltas are actually comparable.
Ask us something harderAnswer
How can you know which chunk was retrieved?
We can't read the model's internals, and we don't claim to. We embed your candidate passages and the competitor's, compare each against the prompt embedding, and cross-reference the cited URL with the fragment and quoted text in the response. That gives a defensible best-candidate attribution — which is enough to act on, because the remediation is the same whether the confidence is 0.8 or 0.95. We report the confidence rather than hiding it.
Ask us something harderAnswer
Is this just SEO with different vocabulary?
There is overlap in the foundations — crawlability, structure, authority — and we'll say so rather than invent a new discipline. The difference is the unit of optimisation. SEO optimises a page for a ranking position. This optimises a passage for retrieval by a system that will never show your page. Different unit, different diagnostics, different fix list, and often different pages entirely: the ones that win citations are rarely the ones that win rankings.
Ask us something harderAnswer
What if we block AI crawlers on purpose?
Then we start with that decision, not around it. Some publishers and IP-sensitive businesses have a legitimate reason to block. We'll model what the block costs you in citation share, show which agents you could allow selectively — retrieval agents that link out versus training crawlers that don't — and let you decide with numbers instead of instinct. If the answer stays 'block everything', the engagement stops there and we say so.
Ask us something harderAnswer
Which models do you cover, and why those?
GPT, Gemini, Claude, Perplexity, Copilot and Google's AI Overviews. Those six cover the overwhelming majority of commercial assistant usage in the markets our clients sell into. We can add others on request, but we won't pad the report with surfaces your buyers don't use just to make the deliverable look larger.
Ask us something harderAnswer
What do you need from our side?
Read access to server logs or your CDN log stream, Search Console, and a technical contact who can confirm rendering behaviour. Content and schema work we can hand over as tickets or implement directly if you give us CMS access. Total time commitment on your side is roughly four hours across the ten days, mostly in the kickoff and the readout.
Ask us something harderStop guessing why the model picked them.
Send your domain and five commercial prompts. We'll run a free 60-prompt pilot across three models and show you your citation share, the competitor holding the slot, and the first three chunks we'd rewrite.
Founded 2014 · 40+ specialists in-house
Works hardest alongside
All AI search services- Parent hubAI Search VisibilityThe full programme across every AI answer surface.
- Sibling serviceGoogle AI OverviewGet cited in the answer box above your own rankings.
- Sibling serviceChatGPT VisibilityBe named when buyers ask an assistant instead of Google.
- DiagnosticAI Visibility AuditFaster, lighter baseline across every assistant.
- FoundationTechnical SEORendering and crawl fixes that citation work depends on.
- AuthorityDigital PRThe third-party corroboration models weight most.
