Use fixed prompts and recorded conditions to measure the mentions and source links that appear in each tested answer. Keep those captures separate from Google Search Console impressions, analytics referral sessions and crawler logs: each describes a different signal, and none provides a universal cross-engine ranking.
A screenshot records one answer; it cannot establish a stable recommendation or citation trend. Product, mode, prompt, account and location can affect the result, so record those conditions with each capture.
Measure the answers you capture, the platform data available to the site owner, and the business outcomes as separate evidence.
Updated 5 November 2026: replaced a retired OpenAI publisher reference and clarified the difference between captured sources, platform impressions and analytics visits.
Separate three types of evidence
AI visibility reports often combine unrelated numbers into one score. Keep three evidence layers separate.
1. Manual answer observations
These show what a product returned under the recorded conditions. Count only mentions and source links visible in the captured answer; a prompt test is a sample, not a universal ranking.
Examples:
- The business was named in the captured answer.
- The website appeared in the product’s visible source links.
- A competitor appeared in the answer; record its position there without treating order as a general ranking.
- The answer stated the wrong city.
2. Platform and website data
These show the platform-reported exposure or visits the available reports can identify. They do not replace manual answer captures.
Examples:
- Google Search Console impressions for AI Overviews and AI Mode.
- Referral sessions identified in web analytics, where source data is available.
- Crawler requests in server logs.
- Landing-page engagement.
3. Business outcomes
These show what happened after visits or enquiries that may be associated with discovery. They do not prove that an AI answer caused the outcome.
Examples:
- Qualified enquiries.
- Calls.
- Proposal requests.
- Assisted conversions.
- Revenue linked to a known lead source.
A crawler request shows access, not an answer mention or citation. A link in one captured answer shows what appeared in that sample, not how many people clicked it. A lead is not revenue.
Build a fixed prompt set
Start with a small fixed set you can repeat each month. For example, choose four to six prompts across discovery, comparison and branded accuracy. Expand the set only when a business decision needs more coverage. A larger set of 20 prompts can help organize broader testing, but it is not a statistically representative measure of all answers.
For a four-prompt baseline, choose three non-branded discovery prompts and one branded accuracy prompt. For a five- or six-prompt set, add another relevant discovery question, a second branded accuracy question or an evidence question if it matters to the business. Keep the same prompts each month so the small sample is comparable.
For broader testing, divide an optional 20-prompt set into:
- Four category prompts.
- Four problem prompts.
- Four comparison prompts.
- Four evidence prompts.
- Four branded accuracy prompts.
Do not place the company name inside discovery prompts. That tests whether the engine selects the business without being told where to look.
Write prompts with enough context to produce a useful answer.
Weak:
Best company near me?
Stronger:
Which Pretoria web design studios publish pricing and let a small-business client own the domain and website files?
The second prompt states the location, category and decision criteria.
Fix the test conditions
Record the conditions with every answer.
| Field | Example |
|---|---|
| Test ID | A unique identifier you assign |
| Date and time | Local date, time and time zone |
| Engine | ChatGPT |
| Product or mode | Search |
| Account state | Logged out |
| Device | Desktop browser |
| Location context | Pretoria, South Africa stated in prompt |
| Exact prompt | Saved without edits |
| Web retrieval | Active, inactive or unclear |
| Clean conversation | Yes or no |
| Answer capture | Screenshot or export path |
| Citation capture | Full URLs |
Use a new conversation for each prompt and record whether Search or web retrieval was active. Prior messages can change the answer.
Repeat the same setup in each cycle where possible. You cannot control every condition, so record any change that may affect the answer.
Start with four core measures
For a first monthly check, record prompt coverage, whether the business was named or explicitly recommended, whether its website appeared in visible source links, and whether the important business facts were accurate. For coverage, write down how many prompts you planned, completed and could not run. Keep the captured answers and exact test conditions with the results.
Count a prompt as completed only when the test ran and you could inspect its answer. Report failed or unavailable prompts separately; do not remove them from the planned total. The formula below uses the original planned count as its denominator.
1. Prompt coverage
Prompt coverage shows how much of the planned test ran successfully.
completed prompts / planned prompts
If some planned prompts fail, report both the number completed and planned. Do not present a partial run as a complete baseline.
2. Mention share
Mention share measures how often the business appears by name in the captured answers. For recommendation prompts, record explicit recommendations separately from simple mentions.
prompts that named the business / completed discovery prompts
Keep branded prompts out of the discovery rate.
3. Citation share
Observed website citation share measures how often the company’s own website appeared in the product’s visible source links for the tested prompts. It describes the captured sample, not every answer the product gave.
completed prompts with a visible link to the company's website / completed prompts in the same prompt set
Record responses with no visible sources as zero visible website citations for that test. If a product or mode does not display sources, keep those tests separate. Track third-party sources separately; a directory can be cited while the company’s website is absent.
4. Accuracy
For each captured answer, score each of these six facts separately:
- Service.
- Location.
- Audience.
- Price or pricing model.
- Founder or team.
- Contact route.
Use the same published business information as your reference for every product and test. Apply these scores:
- 2: correct.
- 1: partly correct or missing a material detail.
- 0: the answer states a fact that conflicts with verified reference information.
If the answer makes no claim about a fact, record Not stated. If the answer makes a claim but your available evidence is insufficient to assess it, record Unverified. Do not score either case as zero; exclude both from the score and denominator. For the Accuracy field, add the points for facts rated 0, 1 or 2, then divide by two times the number of facts scored. Record it as points / available points and show the number of facts scored; the available-points denominator is 2 × facts scored. If no facts can be scored, record N/A (0 facts scored). Compare answers only when the same facts were assessed, or show how many facts were scored for each.
A mention with a wrong service can create low-quality enquiries.
Optional measures for deeper analysis
1. Sentiment and framing
Record whether the answer presents the business positively, neutrally or negatively.
Do not treat every positive adjective as a recommendation. Save the relevant wording and note whether a displayed source supports it.
2. Competitor inclusion
Record every named competitor, the criteria used to include them and where they appeared in the captured answer. Do not call displayed order a general ranking.
This shows which businesses appeared in the repeated sample for a topic, city or proof type.
Do not turn the report into an attack list. Use it to identify missing evidence.
3. Referred sessions
Use analytics traffic-source reports to count AI product sessions when the source is identifiable. These are visits after a click, not a count of answer mentions or citations. If the source is unavailable, record it as unknown or use an optional self-reported source question; do not infer a citation from referral volume.
4. Engaged sessions
A cited page can attract curiosity without commercial fit.
Review:
- Engagement time.
- Pages viewed.
- Scroll depth where available.
- Service-page visits.
- Contact-page visits.
5. Enquiries and assisted conversions
Record form submissions, calls, WhatsApp actions and proposal requests from known or self-reported AI discovery.
Ask:
Where did you first hear about us?
Allow more than one source when the journey involved AI search, Google and a referral.
6. Test volatility
Volatility shows how often the result changes across repeated runs.
repeated prompts whose captured result changed / repeated prompts tested
Repeat the same prompt in the same product, mode and recorded conditions. This describes variation in your sample; it does not establish a business’s global position or a product-wide ranking.
Optional expanded scorecard
Use one row per prompt and product/mode when you need a fuller record. A basic monthly note can use only the planned/completed/not-run prompt counts, test date and conditions, prompt ID, whether the business was named or recommended, whether its website link appeared, and whether key facts were correct.
| Prompt ID | Test status | Product and mode | Mention | Explicit recommendation | Website link shown | Other visible sources | Accuracy (points / available points) | Framing | Competitors named | Attributed visits | Enquiries | Notes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Leave fields blank when a test did not collect the data. Record a zero only when the test ran and the relevant output could be inspected. Keep analytics visits and enquiries in a separate reporting window or source field so they are not mistaken for answer observations.
Measure each engine according to its available data
ChatGPT
Use manual prompt tests for answer behaviour.
Review analytics referral sources for identifiable ChatGPT click-through visits. These count visits after a link click, not answer mentions or citations. Do not assume every ChatGPT visit carries a particular UTM value.
Check whether OAI-SearchBot can access pages intended for search discovery. A crawl request shows access, not inclusion in an answer.
Google AI features and Gemini
Google’s AI Overviews and AI Mode are Google Search features. Standard Search guidance applies; Google does not describe a separate special optimization requirement for these features.
Google Search Console’s generative AI performance report covers impressions from AI Overviews and AI Mode. It can group those impressions by page, country, device and date; it is not a prompt-level citation log and does not report Gemini Apps source links. If the report is unavailable, Google says the property may not meet the report’s impression threshold or may have opted out.
For Gemini Apps, inspect the source links shown in each tested answer and save the answer capture. Gemini does not show source links for every response; if a response has no Sources button, the help page says no links were provided for that response. Keep these observations separate from Google Search Console data.
Perplexity
Use manual prompt tests and capture the visible source links for each answer. Perplexity’s help documentation describes source links in Pro Search answers; record the exact product or mode tested because products can differ.
Check PerplexityBot access and server logs where available. Crawler requests show access, not answer inclusion.
Track Perplexity referral domains in analytics, then connect the landing page to actions.
Diagnose before changing the website
A weak result can have several causes.
| Observation | Likely gap | First check |
|---|---|---|
| No mention and no citation | Access or topic coverage | robots.txt, indexing and source page |
| Competitor cited for your service | Evidence gap | competitor’s cited page and criteria |
| Business named with wrong facts | Entity inconsistency | cited source and public profiles |
| Website cited but business not selected | Commercial clarity | service scope, fit and proof |
| Visits without enquiries | Conversion gap | landing page, call to action and offer |
| Strong branded results only | Discovery gap | non-branded content and corroboration |
| Results change every run | Volatility | test conditions and repeated samples |
Do not prescribe schema for every problem. Fix the observed gap.
Build a monthly report
A basic monthly note can record the four core measures, test conditions, one important limitation and one next action. For a broader review, answer these six questions:
- Which prompts ran?
- Where did the business appear?
- Which sources influenced the answers?
- Which facts were wrong?
- Did anyone visit and enquire?
- What will change next?
An expanded report may also include:
- Coverage by engine, if you need a per-product breakdown.
- Mention and citation share by engine.
- Accuracy issues.
- Competitor patterns.
- New or lost citations.
- Google generative AI visibility where available.
- Referral sessions and engagement.
- Enquiries and lead quality.
- Changes made since the previous cycle.
- Limitations.
Keep the raw answer captures so another person can inspect the claims.
Avoid a single invented AI score
A single composite score can obscure the underlying measures.
If you choose to make a composite score for internal reporting, publish its reproducible formula, weights, data sources and missing-data handling. Show the raw measures beside it, and do not combine crawler access, search impressions, captured answer links, visits and enquiries as though they measure the same outcome.
How often should you test?
A monthly cycle is a practical starting point for a small team that can repeat the same prompt set consistently. Choose a different interval if the business changes more often or cannot review the evidence monthly.
Test more often when:
- The business launches a major service.
- A rebrand changes the entity.
- The website moves domains.
- A public error spreads across sources.
- An engine changes a product or reporting method.
Avoid daily checking. Frequent manual searches encourage teams to react to noise.
What a credible agency should disclose
Ask the provider for:
- The exact prompts.
- The engines and modes.
- The account and location conditions.
- Raw answer captures.
- Citation URLs.
- The scoring formula.
- The difference between observation and analytics.
- Missing data.
- Failed prompts.
- The result the provider cannot guarantee.
Favourable screenshots alone do not show how the full prompt set performed.
Measurement approach
This guide recommends fixed prompts and saved answer captures, supplemented by crawler checks, analytics and enquiry data when the client can provide access. These measures describe different signals; they do not produce a universal ranking or guarantee a recommendation.
The report separates:
- Being crawlable.
- Being cited.
- Being named.
- Being described accurately.
- Receiving a visit.
- Receiving a qualified enquiry.
The useful outcome is a record the business can inspect and a clear next action tied to the evidence.
Read how to improve AI search visibility in South Africa before choosing the work, then review the AI search visibility service for the implementation scope.