Numbers only help you if you know where they came from. This page explains the data sources behind the research, reports and results on this site, how AI search visibility is measured, how I label different kinds of claims, and where the limits are.
Where the data comes from
The research and results on this site draw on a small number of sources, each used for what it’s good at:
| Source | What it’s used for | What to keep in mind |
|---|---|---|
| Google Search Console | Clicks, impressions, average position and queries for sites I have access to | Google anonymises some queries, and average position is averaged across every appearance |
| Google Analytics 4 | Visits, engagement, conversions and traffic from AI assistants | Depends on consent and tracking being set up correctly |
| Google Ads | Search terms, spend and conversions | Only as reliable as the conversion tracking behind it |
| Semrush and Ahrefs | Search volumes, keyword difficulty, competitor visibility | Estimates from panel and clickstream data, not Google’s own figures |
| Screaming Frog crawls | Technical state of a site: status codes, canonicals, links, structure | A snapshot on the day of the crawl |
| Live search results | What actually shows in Google, AI Overviews and AI Mode for a query | Results vary by location, device and personalisation |
| AI assistants | Which businesses and sources ChatGPT, Perplexity, Gemini and Copilot mention | Answers change with wording, model version and time |
Nothing from previous agency employment is used in any study, report or example on this site. Client data only appears with permission.
How results are reported
When I describe a result, I give the totals for a stated period, normally the last 12 months, and name the source. I don’t use a single good week to make a result look bigger, and I don’t compare cherry-picked dates.
For rankings, “top three” means a search term that averages position three or better, with at least 100 impressions and at least one click over the period. That threshold stops terms seen once or twice on a lucky day from inflating the count. “Page one” uses the same rule at position ten.
Results also get context: what else changed on the site in that period, and anything outside my work that could have played a part, such as seasonality or a Google update.
How Search Console data is handled
Search Console is the closest thing to Google’s own record of how a site performs, so most results on this site come from it. A few rules keep the figures honest.
Totals come from the site-level report, not by adding up individual queries, because Google leaves out anonymised queries at query level. Query-level analysis, such as counting top-three terms, uses the full export rather than the few hundred rows the Search Console screen shows. Date ranges are stated exactly, and the most recent two or three days are treated with care, since Google’s figures for them are still being finalised.
When a site has more than one Search Console property, such as separate UK and US stores or a domain property alongside a URL property, I say which one the figures come from and don’t double count.
How AI search visibility is measured
AI answers aren’t stable, so I measure them in a way that can be repeated. For each study or client check, I record:
- the exact prompts or queries used, and how they were chosen
- the platforms checked (Google AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini, Copilot)
- the date and time of collection
- the location and device settings, and whether the browser was signed out
- whether the business was mentioned, whether it was cited with a link, and which other sources were cited
Each prompt is run more than once where possible, because the same question can return a different answer minutes later. Findings are reported as how often something happened across the runs, not as a single screenshot.
For the flagship study, prompts are run through Google AI Overviews, Google AI Mode, ChatGPT search and Perplexity, three times each on separate days, with location set to the UK.
How claims are labelled
Every claim on this site falls into one of six types, and the wording tells you which:
| Type | What it means | Example wording |
|---|---|---|
| Official documentation | Stated by the platform itself | “Google’s documentation says” |
| Observation | Seen in real data from a named site | “In the Search Console data for River Bend” |
| Experiment | A test set up to answer a question | “When I tested this on” |
| Correlation | Two things moving together, not proven cause | “Sites that did X also tended to” |
| Opinion | My view, with reasoning | “My view is” |
| Forecast | A prediction, clearly marked | “My expectation is” |
The difference matters. A correlation isn’t a cause, and a forecast isn’t a fact. When a study only shows correlation, it says so.
How competitors are compared
Competitor comparisons use the same source and the same dates for every site in the comparison. Third-party visibility tools estimate traffic for sites I don’t have access to, so those figures are used to show direction and relative size, not presented as exact traffic.
Competitors are chosen by who actually appears for the searches that matter to the business, not by who the business assumes its competitors are. The two lists are often different, and the difference is usually useful.
Research studies
Every study published in the research section states its collection dates, market, platforms, sample size, how the sample was chosen, what was excluded, how the data was analysed and its limitations. Where it’s possible to share the data, it’s published for download so the study can be checked or repeated.
Studies use data collected specifically for the site. When a study is updated, the page shows what changed and when.
How automated findings are checked
Much of the data gathering on this site and in client work is automated, so the checking has to be deliberate. Before a finding is reported, the underlying data is verified at source:
- Error pages: a page flagged as returning an error is re-requested, because crawlers can be rate-limited and see errors that real visitors don’t
- Duplicates: a URL flagged as duplicate is checked for a canonical tag or noindex before it’s called a problem
- Redirects: a redirect is never mistaken for a canonical, since some tools report one as the other
- Head tags: titles, descriptions and schema are read from the raw HTML, because some tools strip that part of the page
- Report figures: every figure is cross-checked against the platform it came from before the report goes out
It’s slower than trusting the tools, and it’s the reason the tools can be trusted.
Limitations
- Search Console hides some data. Low-volume queries are anonymised, so query-level totals are always lower than site totals.
- Third-party volumes are estimates. Semrush and Ahrefs figures are useful for comparing keywords, not for precise forecasting.
- AI answers move. A result recorded today may not appear tomorrow. That’s why dates, prompts and repeat runs are recorded.
- Personalisation affects what anyone sees. Location, history and device all change results, so collection settings are fixed and stated.
- Correlation isn’t cause. Where a finding could have more than one explanation, the study says so.
Corrections
If you spot an error in a study or a figure, the corrections policy explains how to report it and how the fix is shown. The wider approach to writing and sourcing is in the editorial policy, and the tools that collect much of this data are described on the tech stack page. You can also see how this fits into client work on how I work.