Guides

How to measure AI visibility: impressions, citations, clicks and conversions

Saying that a website “appears in AI” is not enough. To understand whether visibility exists and creates value, crawler access, impressions, citations, real visits and actions on the website must be measured separately.

The short answer

There is no universal counter for AI visibility. Google Search Console measures link impressions within its generative features; Bing Webmaster Tools reports citations and cited pages; analytics only record people who actually reach the website. Credible reporting combines these sources, preserves the meaning of each metric and connects clicks to measurable goals.

Gian Luca Partengo
Web developer and designer since 1995 · Updated: 14 September 2026

AI visibility is a chain of six events, not a score

Each stage answers a different question. Skipping one confuses technical activity, content exposure and commercial outcomes.

01The crawler can reach the page
02The system can store and retrieve it
03A link appears in an AI feature
04The page is displayed as a source
05A person actually visits the website
06The visit produces a useful action

The six stages can influence one another, but none automatically proves the next.

Crawling, impressions, citations, clicks and enquiries are not synonyms

The same page may be crawled without appearing, cited without receiving a click or visited without producing an enquiry.

Stage	What it means	Where it is observed	What it does not prove
Crawl and index	The system can download, process and make a resource retrievable.	Server logs, robots rules, URL inspection and webmaster tools.	That the page was displayed, cited or read by a person.
Impression	A link to the property appears within a generative Google feature.	Generative features performance report in Search Console.	A citation across every AI, a click or a stable position.
Citation	A page is displayed as a source within a measured AI experience.	Bing AI Performance and documented platform checks.	Absolute authority, citation prominence or traffic received.
Click and visit	A person leaves the AI interface and opens a page on the website.	Analytics, referrers and campaign parameters when available.	Every citation that occurred or the quality of the visit.
Conversion	The visit completes a goal: enquiry, purchase, booking or another useful event.	Analytics goals and operational systems connected to the journey.	That every result was caused by the AI platform alone.

The reading rule

Compare metrics from the same stage and platform over time. Do not add Google impressions, Bing citations, manual tests and visits as if they described one observable universe.

Google Search Console now separates impressions in generative features

The dedicated report shows when links to a website appear in AI Overviews and AI Mode without pretending to be a complete report of citations and conversions.

GLOBAL REPORT

Available worldwide since 31 August 2026

Google announced the report on 3 June 2026 and completed its global rollout on 31 August. The documentation states that impressions cover AI Overviews and AI Mode; Search Labs experiments are not included.

The report may not appear when a property lacks sufficient data or access to the feature. The missing panel does not, by itself, prove a technical fault or exclusion from every generative experience.

SCOPE

AI Overviews and AI Mode

The measurement concerns documented generative features within Google Search, not ChatGPT, Copilot or the entire AI market.

DIMENSIONS

Pages, countries, dates and devices

URLs and geographic or temporal trends can be analysed. The current documentation does not provide a Query dimension for this report.

COUNTING

Property or URL aggregation

Two results from the same site in one feature can count as one property impression; when a page is filtered, the count is aggregated by URL.

ABSENCE

No report does not mean no possibility

Low volume can prevent the report from appearing. Accessibility, indexing and available data should be checked before concluding that the website is invisible.

An impression describes link exposure within Google’s scope. It does not say whether the AI preferred that source, how prominently it was presented, whether the user read the answer or opened the website.

Bing AI Performance measures citations; it does not create a new ranking

The public-preview report in Bing Webmaster Tools reveals a stage that was previously almost entirely opaque: pages being used as sources in AI experiences powered by Bing.

01

Total citations

They count how often sources from the website are displayed in measured answers. They do not indicate position, format or prominence.

02

Average cited pages

This shows how many pages contribute on average, but it does not assign an authority score or a role within the answer.

03

Sampled grounding queries

These retrieval phrases are associated with citations and can reveal topics; they remain a sample, not a full list of users’ prompts.

04

Page-level activity

This helps identify which URLs receive citations. It is not a ranking of importance, quality or placement.

05

Trend over time

The time series makes it possible to observe changes and compare them with documented publications, updates and technical work.

06

Declared scope

The report covers Microsoft AI experiences, including Copilot, Bing summaries and selected partners; it does not represent every generated answer on the web.

ChatGPT: separate search eligibility, training controls and referrals

OpenAI explicitly separates the crawler used for search from the crawler that controls possible use of content to improve models.

A public website that does not block OAI-SearchBot can be eligible to appear in ChatGPT Search results and citations. Eligibility does not guarantee inclusion: source selection still depends on the search and retrieval systems.

When a person opens a link from ChatGPT, OpenAI normally adds utm_source=chatgpt.com to the referred URL. Analytics can recognise that click, but the parameter does not reveal answers without clicks or the private conversation that preceded it.

SEARCH

OAI-SearchBot

It controls access for search features and source links. Blocking it reduces eligibility for inclusion in summaries and snippets.

TRAINING

GPTBot is a separate control

The GPTBot directive concerns possible use of content to improve models and does not determine search availability.

VISIT

A referral measures a real click

The recorded session proves that a person reached the site through a link attributed to ChatGPT, not how often the link was displayed.

PRIVACY

Private prompts do not enter analytics

The website owner does not receive the full conversation and cannot automatically reconstruct every question that led to the click.

For crawler configuration that does not confuse search, training and voluntary files, also read llms.txt, robots.txt and AI crawlers.

Plausible measures what people do after the click

The AI Assistants channel automatically groups recognised referrals from ChatGPT, Claude, Perplexity, Gemini, Copilot and other assistants within the normal visitor journey.

01

AI Assistants channel

It groups attributable assistant traffic without requiring a separate report assembled by hand.

02

Specific source

When a referral or parameters are available, it can distinguish which assistant sent the visit.

03

Entry pages

It reveals which content receives clicks and whether AI traffic lands on guides, services, products or unexpected pages.

04

Subsequent journey

The visit can be read alongside subsequent pages to understand whether the person continues, investigates or leaves.

05

Goals and revenue

Events, funnels and revenue connect AI traffic to the actions that genuinely matter to the project.

06

Incomplete attribution

When referrers and UTMs are missing, the visit may appear as Direct / None. Analytics cannot attribute what the browser does not transmit.

To choose a tool based on privacy, consent, speed and depth of data, also read Plausible vs Google Analytics 4.

Each tool observes a different part of the phenomenon

A serious measurement system does not look for software that “sees everything”. It defines what evidence each source can provide and where its reliability ends.

Tool	What it observes	What it proves	Main limitation
Logs and technical tests	Crawler requests, HTTP responses, blocks and served resources.	That a user agent requested a resource and what response it received.	It proves neither a citation, a reading nor a human visit.
Google Search Console	Link impressions in documented generative Google features.	Exposure within the report’s scope and dimensions.	It does not cover other AI systems or currently expose every query.
Bing AI Performance	Citations, pages, trends and sampled grounding queries.	That a website source was displayed in measured experiences.	It does not measure position, citation prominence or clicks.
Web analytics	Attributable visits, entries, navigation, goals and conversions.	That a browser reached the site and completed recorded events.	It cannot see answers without clicks and loses untransmitted referrals.
Prompt monitoring	Answers obtained from a defined set of questions, markets and moments.	That the recorded test produced that result under those conditions.	It is a controlled sample, not real frequency across the universe of prompts.

Some data cannot be shown completely by any dashboard

Transparent reporting also declares what remains outside the measurement. Filling gaps with arbitrary estimates makes a report more impressive but less professional.

CONVERSATIONS

Private prompts remain private

A website does not receive the full text of conversations that cite or mention the business before a possible click.

COVERAGE

There is no universal census

Each platform measures its own scope; external tools query samples rather than every answer produced in the world.

MENTIONS

A brand can appear without a link

An answer may name a business without displaying a clickable source. Analytics and citation reports may not record that exposure.

ATTRIBUTION

Some traffic may appear direct

Browsers, apps, copy and paste and missing parameters can strip referrers and hide the AI origin of a visit.

A declared limitation is part of the data

Good reporting separates measured values, samples, inferences and gaps. It does not relabel direct visits as “probable AI traffic” simply because growth coincides with a publication.

Six errors that distort AI visibility reporting

The most common problem is not a shortage of tools, but using different metrics to support a conclusion that none of them proves.

01

Adding different platforms together

Google impressions and Bing citations have different definitions and may concern the same content during the same period.

02

Calling impressions visits

A displayed link does not prove that a person opened the source or read its content.

03

Treating referrals as total citations

Analytics only see people who click and only when the origin is transmitted in a recognisable form.

04

Confusing bots and audiences

A request from OAI-SearchBot or Bingbot is machine activity; it is not a user, a citation or a conversion.

05

Mixing brand demand and new discovery

Being found by someone who already knows the name and appearing for a generic need represent different opportunities and competition.

06

Judging periods that are too short

Small volumes, updates and rollouts create fluctuations. Without a time series and annotations, noise is easily mistaken for a trend.

A useful monthly dashboard connects exposure, visits and results

A report does not need to be filled with numbers. It needs a small set of defined metrics, comparable periods and a record of the changes actually made.

Area	Metric	Frequency	Decision supported
Technical access	Statuses, blocks, crawler errors and retrievable pages.	Continuous checks; monthly summary.	Remove obstacles before judging visibility.
Google generative	Total impressions plus pages, countries and devices involved.	Monthly with like-for-like comparison.	Identify which areas of the website receive exposure.
Bing and Copilot	Citations, cited pages, trends and sampled topics.	Monthly.	Find content and topics worth strengthening.
AI visits	Attributed visitors, sessions, sources and entry pages.	Monthly and by campaign.	Assess volume and quality of actual traffic.
Actions	Enquiries, purchases, bookings and conversion rate.	Monthly, with journey checks.	Understand whether visits support a concrete goal.
Annotations	New pages, updates, redirects, campaigns and technical changes.	On the day of the change.	Separate plausible correlation from unproven causation.

There is no universal threshold for “good AI visibility”. The useful benchmark is the website’s own time series, segmented by pages and goals and read alongside demand, seasonality and website changes.

An eight-step process for verifiable measurement

Define the scope before collecting data. Constantly changing tools and definitions makes credible comparison impossible.

01
Verify ownership and access
Configure Search Console, Bing Webmaster Tools and analytics correctly for the domain and its actual canonical versions.
02
Check access and indexing
Review robots rules, directives, statuses, canonicals, sitemaps, rendering and relevant crawlers before seeking answers in AI reports.
03
Define website goals
Decide which events represent value and ensure tracking occurs only after the real action without inappropriate personal data.
04
Create a baseline
Record at least one complete period with stable definitions, filters and ranges. Low volumes require longer windows.
05
Segment by page and intent
Separate informational guides, services, products, brand and local demand. Aggregated totals can hide opposing opportunities.
06
Compare platforms without adding them
Place Google impressions, Bing citations and attributed visits side by side while preserving their original labels and units.
07
Read behaviour after the click
Check the entry page, continuation, goal and journey quality instead of stopping at the visit count.
08
Annotate, update and repeat
Record technical and editorial work, wait for a comparable period and see whether the change persists.

Measurement should be designed alongside the website’s structure

In the websites I build, SEO, readability for AI systems and analytics are not three later additions: they must describe the same content and the same user journey.

GLP STANDARD

Make the website verifiable first; interpret the signals second

A page should be public where intended, reachable, semantically clear, linked, canonical and measurable. Only then can impressions, citations and visits be interpreted within a known technical context.

The report keeps verified data, samples and inferences separate. It neither promises citations nor attributes results to a change when the platform cannot prove that relationship.

Readable initial HTML and real links
Consistent URLs, canonicals, hreflang and sitemaps
Relevant crawlers governed deliberately
Useful content, sources and recognisable responsibility
Analytics goals connected to real actions
Regular checks with dates and changes recorded

CONTENT

How to appear in AI

Structure, entities, sources and content that make a page easier to read and verify.

Read the guide to AI visibility

ACCESS

AI crawlers and directives

The actual role of robots.txt, OAI-SearchBot, GPTBot and llms.txt files.

Understand what actually matters

ANALYTICS

Plausible or GA4

Consent, speed, cost and data depth change the measurement system.

Compare the two solutions

PROJECT

The complete technical standard

SEO, GEO, performance, security and measurement belong to the same architecture.

Explore the website standard

Verified primary sources

Definitions and limitations come from the official documentation of the systems described, not commercial estimates from third-party tools.

Google Search Central — Generative AI performance reports
Report announcement, available data and completion of the worldwide rollout on 31 August 2026.

Open source
Google Search Console Help — Generative AI performance report
Scope, dimensions, impression aggregation, Search Labs exclusion and availability conditions.

Open source
Bing Webmaster Blog — AI Performance
Official definitions of citations, cited pages, grounding queries, trends and the report’s limitations.

Open source
OpenAI Help Center — Publisher and developer FAQ
OAI-SearchBot, GPTBot, search eligibility and utm_source=chatgpt.com in referrals.

Open source
Plausible — Track AI traffic
AI Assistants channel, recognised sources, entry pages, journeys and post-click conversions.

Open source
Plausible — Metrics definitions
Definitions of sources, referrals, UTM parameters and Direct / None traffic.

Open source

Sources reviewed and verified on 14 September 2026. Reports, coverage and interfaces may change: always check current official documentation before an analysis.

Frequently asked questions about measuring AI visibility

Direct answers to common questions about impressions, citations, traffic and tools.

How can I tell whether my website appears in AI answers?

Use the official reports available for each platform: Search Console for impressions in Google’s generative features, Bing AI Performance for citations measured by Microsoft, and analytics for visits that actually reached the website. None covers every AI system alone.

Does Google Search Console show queries from AI features?

The generative AI report documentation available in September 2026 lists pages, countries, dates and devices, but not a Query dimension. Exposure can therefore be analysed without reconstructing every question that generated it.

Does an AI citation count as a website visit?

No. A citation means a source was displayed within the measured scope. A visit occurs when a person opens the link and the browser reaches the website; analytics can record it when attribution is transmitted.

Can Plausible recognise traffic from ChatGPT?

Yes, when referrers or parameters allow attribution. Plausible groups ChatGPT and other assistants under AI Assistants and can report entry pages, behaviour and goals. If the referral is missing, some visits may appear as direct.

Are OAI-SearchBot and GPTBot the same thing?

No. OAI-SearchBot concerns search and eligibility for features that link sources; GPTBot concerns possible use of content to improve models. Their directives can be managed separately.

Can I know which prompts led users to my website?

Not completely. Some reports provide dimensions or sampled grounding queries, but private conversations are not delivered to the website. Monitoring tools repeat prompts defined by the owner and produce a sample rather than a census of real questions.

Can Direct traffic come from AI assistants?

It can happen when apps, browsers or copy and paste do not transmit a recognisable referrer, but Direct data cannot prove it. Automatically attributing it to AI would turn a possibility into an unsupported certainty.

What is a good number of AI citations?

There is no universal threshold. Sector, demand, country, website size and platform coverage all change volume. Useful measurement compares like-for-like periods, pages, topics and results against the website’s own baseline.

Is manually checking ChatGPT or Copilot enough?

No. A manual test documents that answer at that time, with that question and context. It is useful for inspecting quality and sources, but does not prove how often a brand or page appears overall.

Does greater AI visibility guarantee more customers?

No. Exposure, citation, click and conversion are separate stages. The website must also provide relevant content, evidence, trust and a clear journey. Measurement reveals where the process stops; it does not create a commercial guarantee.