Methodology

BackTweets does two things, and reports only what is observable - not clicks, not traffic, and not market share. First, it traces a link: paste any URL and we gather, live from open public sources, where it shows up across the web. Second, it measures AI visibility: how often brands, pages and domains get mentioned and cited by AI assistants when they answer real questions.

The link trace is gathered on demand from public data - archive history plus public mentions and citations - and cached briefly. Coverage depends on what those sources have indexed, so a thin trace means we found little there, not that a link has no reach.

AI visibility is drawn from an aggregated view of how entities surface in AI answers across a disclosed set of AI engines, refreshed on a regular cadence. We turn how often something appears in those answers for a topic into a share of voice - always its slice of the sample we can see, never a claim about the whole market. We don’t hand-pick the questions behind the numbers.

AI readability is the other half of AI visibility, and the link report now checks it directly - because “nothing cites this page” and “nothing here is legible” are different problems with opposite answers. We check whether a page’s content is present in the raw HTML (most AI crawlers do not run JavaScript), whether robots.txt admits those crawlers, what structured data is published, and which machine-readable files a site offers. The crawler, structured-data and file checks come from Geordy, built by the same team as BackTweets; the raw-HTML check is ours, and each finding is labelled with its source.

Those are findings, not a score. We do not grade a site, and nothing in that section predicts whether AI answers will cite it - being readable is a precondition, not a promise. Where a check cannot complete we say it is unknown rather than reporting it as passing.

Every figure is a sample, not a census. No complete record exists of everything everyone asks an AI, so we treat each number as a disclosed sample, show the coverage and sample size behind it, and never imply total search volume or market share. When the engines or regions we cover change, we say so rather than splicing incompatible periods together - so week-to-week movement reflects real change in AI answers, not a change in how we measure.