Observe, do not assume
Answers vary. A single manual query is an anecdote, so agreed prompts are observed repeatedly and time-stamped.
The Citemark evidence method
Five gates between a website and a defensible AI recommendation.
We measure what appears, inspect why it may be happening, improve the evidence we can control, and re-run the same agreed questions. The method is published because a client should be able to audit the auditor.
01EligibilityFOUNDATION
02Entity clarityFOUNDATION
03Answer coverageOWNED EVIDENCE
04Extractable evidenceOWNED EVIDENCE
05CorroborationINDEPENDENT
Separate access from visibility
A readable site is eligible, not guaranteed to appear. The free scan tests readiness; paid measurement tests observed answers.
Improve evidence, not “the model”
We change public pages, structured facts and source consistency. We do not claim access to an engine’s ranking system.
The five-gate assessment
Every layer ends in evidence and a pass condition.
The order prevents waste. Publishing ten new pages cannot solve a site that is inaccessible; schema cannot rescue claims that are absent or inconsistent; owned copy cannot substitute for independent trust.
LAYER 01
Eligibility
Can retrieval systems access and use the page?WHAT WE EXAMINE
Rendering, indexability, robots and CDN rules, internal discovery, snippet controls, status codes and canonical signals.
CLIENT EVIDENCE
Crawler access matrix · rendered-text check · indexed URL sample
PASS: Priority pages return usable public text and are eligible for retrieval.LAYER 02
Entity clarity
Can the business and its claims be identified consistently?WHAT WE EXAMINE
Canonical name, address, service area, people, credentials, profiles and visible facts aligned with accurate structured data.
CLIENT EVIDENCE
Entity fact sheet · profile discrepancy log · schema validation
PASS: Important facts are consistent, visible and attributable to one clear entity.LAYER 03
Answer coverage
Do priority pages resolve the buyer’s real decision questions?WHAT WE EXAMINE
Decision-stage question research, page mapping, pricing and fit content, comparisons, FAQs, original detail and internal linking.
CLIENT EVIDENCE
Prompt-to-page map · content brief · published answer assets
PASS: Each priority question has a useful, specific page or section—not thin copy.LAYER 04
Extractable evidence
Is there specific evidence a system can quote or verify?WHAT WE EXAMINE
Named expertise, methodology, fees or ranges, constraints, dates, original facts, service definitions, proof and source attribution.
CLIENT EVIDENCE
Claim register · reviewer/sign-off trail · evidence blocks
PASS: Commercial claims have a visible basis and do not depend on unsupported superlatives.LAYER 05
Corroboration
Do independent sources support the same facts?WHAT WE EXAMINE
Review platforms, regulator records, association and partner profiles, directories, editorial opportunities and citation-source gaps.
CLIENT EVIDENCE
Source graph · submission log · outreach record
PASS: Owned claims and credible third-party records agree. Placement is never guaranteed.Measurement protocol
A baseline that can be repeated—not a magic visibility score.
AI answers are stochastic, location-sensitive and subject to product changes. We use a fixed, documented sample so month-to-month movement is interpretable without pretending it represents every possible user.
EXAMPLE OBSERVATION MATRIX15 prompts × 3 runs
named cited absent
Agree the buyer questions
Discovery, comparison, fit, trust and price prompts are tailored to the sector, offer and locations.
Name the surfaces
The consumer products available for the engagement are stated before work starts. Surface and location context are recorded.
Capture the answer
Each prompt is observed three times. Brand mention, position, wording, citations and named competitors are retained.
Report with limits
We show the numerator, denominator, date and sample. We do not extrapolate the set into a market-wide traffic claim.
Evidence boundaries
What each product can—and cannot—tell you.
Public-site readiness
Checks access, visible entity facts, answer structure and surface-level trust signals.
Does not test live recommendation frequency.Observed answer evidence
Tests an agreed prompt sample and records mentions, citations, competitors and gaps.
Does not predict future answers or leads.Implemented improvements
Changes owned pages and facts, maintains the evidence set and pursues defined corroboration work.
Does not buy or guarantee placement.What we deliberately avoid
No theatre disguised as optimisation.
Special-file claims
Google says no special AI text file or schema is required for AI Overviews or AI Mode. We do not sell an llms.txt file as a strategy.
Bulk “AI content”
Google recommends people-first, original, useful content and warns against scaled material made mainly to manipulate search.
Hidden prompt tricks
Hidden text, prompt injection and machine-only claims create risk without producing defensible public evidence.
Guaranteed citations
Eligibility, indexing, rich results and generated answers remain at each platform’s discretion.
Primary sources
The parts platforms publish themselves.
Our operating rules start with first-party documentation, then use observed prompt and citation evidence to prioritise the client-specific work.
See your first layer
Start with a real inspection of the public site.
The free scan checks whether the foundations are usable. The £195 Snapshot adds repeated prompt evidence when you need to know what appears.