By Adriaan Brits, CEO of Sitetrail
Someone on the team asks ChatGPT a question, sees the brand recommended, and shares a screenshot in the group chat. A week later a different colleague asks something similar in Perplexity and the brand isn’t mentioned at all. Neither screenshot tells you much on its own. AI search answers change with the wording of the question, the engine, the date, the user’s location and history, and plain randomness. If you want to know whether your visibility in AI search is actually improving, you need a record you can repeat and audit.
This article sets out a lightweight method I call an AI citation evidence log. It works for Google AI Overviews and AI Mode, ChatGPT, Perplexity and Gemini, and it needs nothing more than a spreadsheet and some discipline.
One thing to keep in mind: AI increasingly provides personalized answers. Many “AI SEO tools” don’t account for this fact.
Why screenshots are not measurement
A single favourable answer is an anecdote. Generative engines don’t return a fixed ranked list the way a classic results page roughly does. The same prompt can give different wording, different sources and a different set of brands from one run to the next. Results also differ between logged-in and logged-out sessions, between countries, and between model versions, and the providers change those versions without notice.
That doesn’t make the channel unmeasurable. It means you should measure it the way you’d measure any noisy system: keep the inputs fixed, sample repeatedly, write everything down, and look at trends rather than single observations.
Step 1: Build a fixed prompt set
Start with 20 to 40 prompts that reflect how real customers ask about your category. Group them by intent:
-
Category discovery: “What are good tools for tracking delivery fleets?”
-
Comparison: “X vs Y for a small logistics company”
-
Problem-led: “How do I reduce failed deliveries?”
-
Brand-specific: “Is [Brand] reliable?” and “What does [Brand] cost?”
-
Local or segment variants, where they matter to the business
Write each prompt down exactly and give it an ID. Don’t reword prompts between rounds. If you discover a better phrasing, add it as a new prompt with a new ID and keep the old one, so your history stays comparable.
Include prompts where you currently don’t appear. A log that only tracks your wins measures your optimism, not your visibility.
Step 2: Set a recording cadence and conditions
Pick a cadence you can keep up with. Fortnightly or monthly suits most teams. Weekly makes sense during a launch or after significant site changes. For every round:
-
Use the same engines, and note the mode (for example AI Overviews vs AI Mode, or ChatGPT with search on).
-
Use clean sessions where possible: logged out or a dedicated test account, with history and memory off where the product allows it.
-
Record the location and language setting.
-
Run each prompt more than once. Two or three runs per prompt per engine is a practical minimum, so you can tell a stable result from a lucky one.
Step 3: Capture the same fields every time
Consistency matters more than detail. Here’s a simple template:
|
Field |
Example |
Notes |
|---|---|---|
|
Date / time |
2026-10-16 10:30 |
Include the time zone |
|
Prompt ID |
P07 |
From your fixed set |
|
Engine + mode |
Perplexity (default) |
Note the model or version if shown |
|
Run # |
2 of 3 |
Repeat runs per round |
|
Session conditions |
Logged out, UK, EN |
Location, account state |
|
Brand present? |
Yes / No |
Anywhere in the answer |
|
Mention type |
Linked citation / unlinked mention / both |
A source link vs a name in the text |
|
Cited URL(s) |
Exact pages, yours and third-party |
|
|
Position |
1st of 4 brands named |
Order in the list or answer |
|
Sentiment / framing |
Neutral, “budget option” |
How you’re described, quoted briefly |
|
Accuracy |
Price outdated |
Factual errors about you |
|
Competitors named |
Brand A, Brand B |
Who appears instead of or alongside you |
|
Evidence |
Screenshot / share link |
Stored with a filename tied to the row |
Two of these fields deserve extra care.
Mention vs link. Being named in the answer text and being cited as a source are different results. You can be recommended with no link at all, or your page can be cited as a source while a competitor gets the recommendation. Log them separately.
Cited URLs, including third-party ones. These are often the most useful data in the log. If the engines keep citing a particular review site, directory, forum thread or news article when they answer questions in your category, that tells you where the evidence they rely on actually lives.
Step 4: Summarise without overclaiming
After each round, calculate a few simple figures per engine and per intent group:
-
Presence rate: the share of runs in which the brand appears at all.
-
Citation rate: the share of runs in which one of your URLs is cited.
-
Consistency: for each prompt, whether you appeared in all runs, some runs or none.
-
Competitor share: how often each named competitor appears across the same runs.
-
Accuracy issues: a running list of errors and where they seem to come from.
Report these as ranges and trends over several rounds, with the sample size beside every figure. “Present in 9 of 24 runs on Perplexity, up from 4 of 24 last month” is honest and useful. “We rank #1 in ChatGPT” is neither.
Step 5: Turn findings into entity and evidence fixes
The log is only worth keeping if it changes what you do. Common patterns, and the fixes that follow from them:
The engine misdescribes what you do. That usually points to an entity problem: your name, category and offering aren’t described clearly and consistently across the web. Make sure your About page, product pages, organisation structured data, business profiles and major directory listings all describe you the same way. Correct outdated third-party descriptions where you can.
Competitors are cited from sources where you’re absent. Look at the third-party URLs in your log. If a particular comparison article, industry list or community discussion keeps being used, ask whether you have a legitimate reason to be in it. Earned coverage, genuine expert commentary and accurate listings help. Unlinked brand mentions in credible places can matter too, because the engine is building a picture of who you are, not just counting links.
You’re mentioned but never cited. The engine may know your name but find nothing on your site that answers the question directly. Publish pages that state facts plainly: pricing approach, specifications, methodology, original data and named authors with real experience. Content that contains verifiable evidence is easier to cite than content that only makes claims.
Facts are wrong or outdated. Update the source of truth on your site first, then the third-party pages the engines are citing. Re-check those prompts in the next round.
Google’s own guidance on AI features and your website is worth reading here. In short, the fundamentals that make pages eligible for normal search results also apply to AI features, and there’s no special markup that guarantees inclusion.
Step 6: Be honest about the limits
Any report built on this log should say plainly what it can’t show:
-
Personalisation. Your test sessions aren’t your customers’ sessions. Account history, memory, location and device all shape answers.
-
Volatility. Models and retrieval systems change without announcement. A drop can come from a model update rather than anything you did.
-
Sampling. A few dozen prompts run a few times each is a sample, not a census. It can show direction, not precise market share.
-
Attribution. Most AI answers send little traffic you can measure, and analytics can’t always tell you that a visit started in an AI answer. Compare the log with branded search demand, direct traffic and sales conversations, but don’t claim a causal link the data can’t support.
-
Non-repeatability. A single favourable answer, however good it looks, isn’t a result you can promise to a client or a board. The value of the log is that it replaces those anecdotes with a repeatable record.
A minimal starting routine
If this seems like a lot, start small: 20 prompts, three engines, two runs each, once a month. That’s 120 observations per round, which a person can log in an afternoon. After three rounds you’ll have a baseline, a list of the sources the engines trust in your category, and a short, specific list of entity and evidence fixes. That will tell you far more than any screenshot.
About the author: Adriaan Brits is the CEO of Sitetrail, a digital marketing and PR company, and holds an MSc in Marketing. Disclosure: the author leads Sitetrail, which offers SEO and PR services. This framework draws on the Generative Engine Optimization module of Sitetrail’s free SEO course. No purchase is required to use the method described here.