How to measure GEO results: mentions, citations, visits and enquiries

To measure GEO results, keep four records separate: what AI answers say about the brand, which pages they cite, which identifiable visits reach the website, and which qualified outcomes follow. Compare the same relevant question groups over time. Do not compress all four into a score that hides their different meanings.

There is no complete public record of every buyer's AI conversation. A prompt panel is a sample, and analytics only captures what its collection and attribution allow. Ahrefs recommends examining groups of prompts and several sources of evidence, rather than treating one answer as a stable ranking. The method below is a practical observation design, not a claim that GEO impact can be measured perfectly.

Original diagram. These are separate records that may be connected where evidence permits.

Start with a baseline before changing the content

Record the current public URLs, their content versions, access state and publication dates. Identify the exact change you intend to make. If the website is also undergoing a redesign, new advertising campaign or major product launch, note those events: they can affect discovery and enquiries at the same time.

Then record existing search performance, identifiable AI referrals and qualified enquiries for a useful comparison window. A small business with few enquiries may need a longer window to interpret changes. Choose the window based on the business cycle and available data; do not promise improvement after a fixed number of days.

The baseline can contain zeros. It can also contain unknowns. “No citation found in the sampled answers” is a valid observation. “AI never cites this website” is a much broader claim that the sample cannot establish.

Choose questions that represent decisions

Build question groups from customer questions, sales calls, search data and relevant public discussions. Mark generated expansions as hypotheses. Separate branded questions, such as “What does this supplier deliver?”, from non-branded choices, such as “How do I choose a supplier for this application?” A branded question can be useful for checking accuracy without showing competitive discovery.

For an ecommerce store, useful groups may cover compatibility, alternatives and the first use of a product. For a manufacturer, group questions by application, specification and supplier qualification. Avoid filling the panel with slight rewrites of one favourable branded question.

Save the wording and keep a stable core set for comparison. Add new questions in a separate group so that a changing denominator does not look like a sudden performance shift. A team's own question set is not proof of the market's actual prompt frequency; measured demand needs a named data source and its limitations.

Keep the conditions comparable

For each observation, record the platform or interface, model if shown, whether web search was active, language, market or location setting, date and time, conversation context and any significant account conditions. Start a fresh conversation where practical. Personalisation and changing retrieval can still affect the answer, so describe the method rather than claiming perfect control.

Save the full answer and the citation URLs. A screenshot can help verify the displayed state, but the text and links make later review easier. If an answer does not display sources, record that absence instead of inferring an unseen source from similar wording.

Repeat the same core panel at a chosen cadence that the team can sustain. Inspect variability across observations and platforms. Repeating a prompt ten times until it produces a preferred answer and reporting only that answer creates a misleading result.

Use definitions that someone else can reproduce

Use a denominator. If 6 of 24 sampled answers name the brand, report 6/24 sampled answers, with the question set and conditions. Do not label it “25% of all AI searches”. When comparing competitors, use the same panel and naming rules; a different sample cannot support a direct share comparison.

A citation record should distinguish your own URL from a third-party page mentioning your brand. Keep the exact URL, the source title and the passage or context it supports. Counting every link anywhere in the interface as an endorsement will inflate the result.

Read an example without mistaking it for proof

The following is fictional demonstration data, not MadeInRed performance or a customer case study. It shows how a team might interpret a small panel.

The table does not establish that the content change caused the enquiry. The sample is small, time periods can differ and other activity may contribute. A useful report would attach the observations, show the publication event and note competing explanations. It would not turn these four rows into a fabricated percentage uplift claim.

Suggested observation sequence. Timing should follow the business and available evidence, rather than a guaranteed result date.

Connect website analytics without overstating attribution

Check identifiable source/medium and landing-page records in the analytics system you actually use. Keep the date range, filters and channel definitions alongside exports. Verify that the relevant measurement and consent configuration works; an empty report can reflect collection limits as well as an absence of visits.

For Google Search's AI features, Google states that their traffic is included in the overall Web search performance data in Search Console. Do not describe that aggregate as a separate complete ChatGPT citation report. Search Console, web analytics, server logs and saved AI answers answer different questions.

An optional “How did you hear about us?” field can add qualitative information. Preserve the customer's wording and avoid forcing every answer into a last-click model. For B2B, connect the enquiry to a qualification record or CRM stage only when the relevant permissions and evidence exist. A page visit alone is not pipeline.

Decide what to change from the pattern

If the page cannot be accessed or its important text is missing, fix that concrete issue. If competing sources answer a useful buyer question and yours does not, improve the answer with relevant evidence. If the brand is mentioned incorrectly, identify the inaccurate fact and the sources available to support a correction.

If citations appear but few identifiable visitors arrive, inspect whether the answer already satisfies the user and whether the linked page offers a useful next step. If visits arrive without qualified enquiries, review audience fit, the page's promise and the enquiry path. Do not assume that more citations will repair a poor commercial fit.

If proposed articles repeatedly cover an existing intent, use the content library to improve the existing asset or research a distinct question. Measurement should guide the next useful action. It should not manufacture a favourable headline from whichever number moved.

A compact monthly record

Keep the question set version, observation conditions, saved answers, exact citations, publication changes, analytics exports, qualification definitions and limitations together. Write three sentences: what changed in the evidence, what remains uncertain, and what action the team will take next.

This record makes GEO results reviewable. It also makes an honest “not enough evidence yet” possible without denying that useful content has already been delivered. For the writing work that precedes measurement, see GEO content optimization.

Sources

• Ahrefs — How to track AI visibility
• Ahrefs — How to choose prompts
• Google Search Central — AI features and your website

Read 30 October 2026. Inline links identify the claims they support. Product descriptions are first-party claims, not independent trial results.