Guide · SEO and AI visibility
How to measure AI citations and visits.
A citation, a visit and a business enquiry are different outcomes. Record each separately, with its URL, period and data source. This workflow helps you assess which articles support discovery without turning a few answer screenshots into a universal AI ranking.
Define what counts as a citation
A mention names your brand. A citation identifies a page as a source. A recommendation proposes your service. During manual review, save the answer and destination URL: seeing your name does not establish that the system used your article.
Decide what you want to learn from the measurement. For example, does a guide about integrations help readers understand when to ask for development support? Review citations in that context, alongside visits and enquiries.
Read each report within its scope
Search Console’s generative report shows impressions by page, country, device and date; it may be absent when data is insufficient. Bing AI Performance reports citations and a sample of retrieval queries across supported experiences, not all AI platforms.
Save the report name, filters and dates with each export. Do not add unlike units together or add a subset to a total that already contains it.
Connect referrals to landing pages
Review referring domains and landing pages in your analytics. An identifiable assistant referral establishes a visit, not the full number of citations. Copied links and applications can lose referral information.
This website uses Cloudflare Web Analytics with a production-host filter. Preview visits are excluded; our own production visits can remain. Keep sampling and missing dimensions visible alongside the result.
Create a repeatable question sample
Choose diagnostic, comparison and purchase questions a customer might ask. Include language and buying context without placing your brand in every prompt. Keep one fixed comparison set and a separate set for exploring new questions.
Hypothetical example: “Can a bot handle repair requests?” and “Can I automate work in my old software?” concern the same problem but are different tests. Record platform, search mode, date, exact prompt, answer, cited URL and attribution errors. Repeat observations; a single answer is unstable.
For each saved answer, mark the URL as a source link, an unlinked mention or no appearance. If the answer recommends a competitor but cites your technical explanation, record both facts: source use and business recommendation are separate observations.
| What it establishes | What it does not establish | |
|---|---|---|
| Cited page | An observed answer links to a specific URL. | That the reader visited or endorsed the business. |
| Referral visit | A recorded visit arrived with an identifiable source. | Every citation or every visit from that assistant. |
| Qualified enquiry | Someone described a relevant project. | Which earlier citation influenced them without further evidence. |
Turn a finding into a specific correction
If a sentence is cited out of context, check whether its condition or limitation is close enough. If a buying question leads to a general guide, add a relevant next step. If there are no citations, investigate access and content before assuming a penalty.
Use a record containing the question, intended page, observed evidence, missing distinction, change and review date. Do not publish another article for every wording variation when one existing section can answer it.
If source access is missing, start with the indexing diagnosis.
Compare periods and preserve uncertainty
As an editorial routine, review after 30, 60 and 90 days. Compare equivalent periods and separate languages, countries and new pages from established ones. These are working intervals, not promised growth deadlines.
Report four columns: observed exposure, identifiable visits, enquiries received and attribution limits. Do not calculate a conversion rate from citations in one report and contacts in another without a verifiable connection. Mark the outcome “not attributable” when that connection is missing.
Sources and further reading
Questions before we start
Can I measure every ChatGPT or Perplexity citation?
Not with this workflow. Referrals show identifiable visits and manual reviews observe a sample of answers. Neither is a complete citation census.
Does a citation without a click have value?
It may provide exposure, but it does not establish a visit or sale. Record it as exposure and assess business outcomes separately.
Do I need a paid GEO tool?
Not to start. Define the questions, URLs and decisions you want to observe first. A tool is useful if it improves collection and review with clear limitations.
Which questions should I track for my business?
Start with questions raised before an enquiry: cost drivers, alternatives, compatibility and what could go wrong. Assign each to the page that should answer it. Keep branded searches separate from questions where the customer does not yet know you.