To measure AI search visibility, use a stable research protocol and keep it separate from website analytics. A citation observed in a chatbot answer is evidence of that answer. It does not reveal total impressions, prove a ranking position, or establish that the viewer visited your website.
The protocol below is a suggested measurement method for the B2B AI search visibility system. Use it to make comparisons more interpretable, not to manufacture a universal visibility score.
Build prompts around real buyer decisions
Choose questions about a service category, a comparison, an implementation concern, and a risk. Avoid filling the set with brand-name prompts that already instruct the system to discuss your company.
For a reporting consultancy, useful questions might concern combining advertising and CRM data, validating totals, or deciding whether to build internally. Label branded and unbranded questions separately so they do not inflate each other’s results.
Keep the observation conditions consistent
Record the platform, mode, account context, language, locale when controllable, exact prompt, and time. Use a fresh conversation for each question if that is your chosen protocol. Preserve the same conditions at the next review wherever possible.
Run repeated observations rather than trusting one response. The number of runs is a research design choice, not a platform requirement. Report both the sample size and the variation you observed.
Define the outcomes before counting
| Outcome | Suggested definition |
|---|---|
| Brand mention | The answer names the business |
| Citation | The answer links to an identifiable business page |
| Relevant citation | The linked page supports the buyer question |
| Referral | A website visit has identifiable source information |
| Qualified enquiry | An enquiry meets your documented qualification rule |
Keep the cited URL, surrounding context, and a saved observation. If an answer cites your homepage but makes an unsupported service claim, record that as an issue rather than treating it as an unqualified success.
Calculate a transparent sample rate
One useful measure is observed citation rate: runs citing an in-scope page divided by eligible runs. Label it as a sample measure and keep branded questions separate. Document failures and unavailable results so reviewers understand the denominator.
Do not combine different platform rates into a single market-share claim. A change in the prompt mix can change the rate even when the underlying content has not changed.
Track traffic independently
OpenAI’s publisher FAQ describes identifiable ChatGPT referral tracking, including its UTM source parameter. Preserve available attribution rather than assuming every AI-origin visit is invisible.
Other visits may lack clear source information. Keep those classified according to the available evidence. A sudden rise in direct traffic does not by itself demonstrate AI discovery, and observed citations do not tell you how many people clicked.
Explain changes before celebrating them
Keep a release log covering new pages, changed evidence, technical fixes, and revisions to the prompt set. Compare stable cohorts. If one platform improves while another declines, report that variation instead of smoothing it into a reassuring average.
Also review the commercial result. More relevant citations with no enquiries may indicate weak offers, unclear next steps, or a research audience that is not ready to buy. Investigate the journey rather than rewriting every article.
Use the findings to improve specific pages
When a question consistently lacks a useful answer from your site, inspect the comparison page, claim evidence, and access checks associated with it. Change the part that is weak, then repeat the established protocol.
This creates a defensible learning cycle: publish, observe, inspect, improve, and retest.