Measuring whether AI answers cite you
This is a measurement service and not a way to earn citations. What it establishes is whether a set of queries produces answers that reference your site — and what the instrument can and cannot tell you.
AI answers are a discovery surface with no reporting
Search has a console that reports impressions, clicks and the queries behind them. AI answers have nothing equivalent: a citation appears in a generated response, varies between sessions and users, and leaves no record unless somebody looks. So a site can be cited regularly and not know it, or be losing citations it used to have and not notice — and neither is visible in any tool the site already has.
- It is not known whether AI answers cite the site at all
- Traffic from AI surfaces appears in analytics with no way to see what drove it
- The site is cited in some contexts and not others with no established pattern
- Competitors appear in answers where the site does not
- A change was made for AI visibility and there is no way to tell whether it worked
- Nobody can say which pages are being cited, or for what
- The question comes up in every review and there is no answer to give
- Reporting covers search and not this, so it is discussed anecdotally
The people who usually bring us this problem
A team asked whether they are cited
The question comes up regularly and there is no measurement behind the answer.
A team that has done AI-visibility work
Changes were made and there is no way to establish whether they had an effect.
A team that needs this in reporting
AI surfaces are a real source of discovery and they are absent from every report.
What this costs while it goes unfixed
Engineering faults are rarely confined to the engineering layer. These are the commercial consequences we see most often.
The instrument has limits and they have to be stated
AI answers vary between sessions, between users and over time, and the systems behind them change without notice. A measurement is a sample of queries at a point in time, repeated on an interval — it establishes a trend and a proportion, and it does not establish what any individual user will see. A monitoring service that does not say this is presenting a sample as a fact.
A citation is not a click and not a ranking
Being referenced in a generated answer is a different thing from being linked and a different thing again from appearing in a list of results. The measurement has to distinguish them — a citation, a link, a mention — because the actions they suggest are different, and a report that counts them together answers a question nobody asked.
Detection is attribution and it is where the errors are
Establishing that an answer referenced your site, as opposed to a site with a similar name, a syndicated copy, or a page quoting you. Attribution rules decide what the numbers mean, and a service that does not describe them is producing figures whose basis cannot be examined.
Nothing here produces a citation
Measurement establishes what happened. What changes it is the work on the site — crawlability, entity consistency, structured data, content — and that is the other page. A monitoring service that implies otherwise is selling the reporting as though it were the mechanism, which is the failure mode this page's gate explicitly guards against.
Capabilities
Each of these is work we carry out, not an area we advise on.
Query set design
The questions the measurement asks, derived from what the site is about and what its audience would ask rather than from a keyword list. The set is the instrument, and it is reviewed periodically because the questions people ask a generative system are not the ones they typed into a search box.
What is measured per query
Whether the site is referenced, where in the answer, in what form — a link, a named mention, a quotation — and which URL was referenced. Each is recorded separately because they mean different things, and collapsing them produces a number that cannot be acted on.
Attribution rules, stated
How a reference is established and how a lookalike is excluded, written down and applied consistently. The rules are the measurement's basis, and stating them is what allows a figure to be examined rather than taken on trust.
Repeated measurement on an interval
The same query set at a defined frequency, so the result is a trend rather than a snapshot. A single measurement of a surface that varies between sessions establishes very little; the interval is what turns it into something usable.
Competitor and share-of-answer context
Which other sites appear in the same answers, so a citation is reported as a position within an answer rather than in isolation. Being cited alongside four competitors is a different situation from being the only reference.
Change correlation
Recording when the site changed and what it changed, against the measurement series. It does not establish causation — nothing about this surface does — but a change that coincides with a movement is worth recording, and the alternative is that nobody can remember when anything happened.
Reporting that states its own limits
The interval, the query set, the number of samples and the variance, alongside the findings. A report that presents a proportion without its basis invites conclusions the measurement cannot support, and on a surface this variable that is the main risk.
The boundary, made explicit
What this establishes and what it does not, in the report itself. It is what stops the measurement being read as a promise about future citations, and it is the reason the service can be described honestly at all.
Engineering methodology
The sequence is deliberate. The order is usually what determines whether the work holds or has to be repeated.
Build the query set from the site's subject, not from keywords
The questions a person would ask a generative system about this subject. Search queries and conversational questions are not the same thing, and a set built from a keyword tool measures a surface that is not the one being measured.
State the attribution rules before the first measurement
How a reference is identified, how a lookalike excluded, how a syndicated copy treated. The rules determine what every later figure means, and writing them first is what makes the series consistent rather than re-interpreted each time.
Measure repeatedly, and report the variance
The same set on a fixed interval, with the spread across samples shown. On a surface that varies between sessions, a single figure is close to meaningless and a series with its variance stated is usable.
Separate a citation from a link from a mention
And report them separately, because the actions they suggest differ. Collapsing them produces a total that cannot be acted on and a report that answers a question nobody asked.
Record site changes alongside the series
What changed and when, so a movement can be placed against it. This does not establish causation, and it is the difference between a report that can be reasoned about and one where nobody can remember what happened when.
Report what the instrument cannot establish
In the report rather than in a caveat nobody reads. The system changes, the samples are a sample, and no measurement produces a citation — all three stated alongside the findings, because the conclusions drawn from a report are the risk this work carries.
What an engagement produces
Documentation is a deliverable, not an afterthought. On most of these engagements a large part of the value is a defect report precise enough for another team to act on.
Setup
- The query set, with its basis
- What is recorded per query: reference, position, form, URL
- Attribution rules, written down
- Measurement interval and sample count
- Which other sites are tracked for context
Recurring measurement
- The query set measured on the defined interval
- Citations, links and mentions recorded separately
- Competitor presence in the same answers
- Site changes recorded against the series
- Variance reported alongside the findings
Reporting
- The trend, with its basis and its limits
- Which pages are referenced, and for which questions
- Where the site appears alongside competitors, and where it does not
- What changed on the site, and what the series did afterwards
- What this measurement does not establish, stated plainly
Architecture and technology
What the measurement records
- Whether the site is referenced in an answer to a given question
- The form: a link, a named mention, or a quotation
- Where in the answer the reference appears
- Which URL was referenced
- Which other sites appeared in the same answer
- How the result varies across repeated samples
What it cannot establish
- What any individual user will see, since answers vary per session
- Whether a citation produced a visit, which is a different measurement
- Why a change in the series happened
- What the systems behind the answers will do next
- That any action will produce a citation — nothing does
- A ranking, because there is no ranked list to be positioned in
If this is not quite your problem
These overlap at the edges. Sending you to the right page is more useful than having you work it out.
You want the work rather than the measurement
The discipline: crawlability, entity consistency and machine-readable facts.
AI search optimisationThe entities need describing consistently
One identity per entity, referenced identically everywhere.
Structured data implementationThe content has to be crawlable at all
What is crawlable, indexable and reachable, and what limits each.
Crawl and indexationYou want the whole position established once
An audit delivered as a prioritised engineering backlog.
Technical SEO auditFrequently asked
How is this different from your AI search optimisation page?
Direction. AI search optimisation is the work — crawlability, consistent entities, machine-readable facts, content that answers a question directly — and it is the page for someone who wants to change something. This is the measurement of whether any of it had an effect, and it is the page for someone who needs to know what is happening. They are bought for different reasons and they are frequently bought together, in the order that makes sense for the situation: measure first if the position is unknown, work first if it is not.
Will this get us cited more?
No, and no measurement service can. What this establishes is whether a defined set of questions produces answers that reference your site, how that changes over time, and where you appear relative to competitors. What changes it is work on the site, which is the other page. A monitoring service that implies it produces citations is selling reporting as though it were the mechanism, and that is the specific thing this page is built not to do.
How reliable is the measurement?
It is a sample, and its reliability is stated rather than implied. Answers vary between sessions and users, the systems behind them change without notice, and a single measurement establishes very little — which is why the same query set is measured repeatedly and the variance is reported alongside the findings. What the series gives you is a trend and a proportion with its basis visible. What it does not give you is what any individual user will see, and the report says so.
How do you know a citation is actually us?
By attribution rules that are written down and applied consistently — which is the part of this work most likely to be done badly and least likely to be visible. A site with a similar name, a syndicated copy of your content, or a page quoting you can all look like a citation if the rule is loose, and the resulting figures would be wrong in a direction that flatters. The rules are part of the setup deliverable so that a figure can be examined rather than taken on trust.
Can you tell whether a citation sent us traffic?
Not from this measurement. Whether a citation produced a visit is a different question with a different instrument — referrer data and analytics, which on these surfaces is frequently unattributed because the visit may not carry a usable referrer. So the two are kept separate: this measures whether the reference exists, and traffic attribution is a separate thing that is usually partial. Conflating them would produce a number that appears to answer the commercial question and does not.
Related capabilities and work
Bring us the problem you have not been able to fix
Describe what is happening rather than what you think the cause is. If we are not the right people for it, we will say so.