When a vendor reports that some change tripled their AI citations, the right first question isn't "how?". It's "compared to what, and how do they know it was the change?" A useful worked example landed in June 2026: a 19-week dataset, published by the brand-visibility platform waikay.io about its own site, reporting weekly Bing AI citations rising from a 167 average to 614. About a 3.7× lift: after it added entity markup. It's a genuinely interesting result. It's also a textbook case for reading AI-visibility studies critically.
This is part of the measuring AI visibility series, and a companion to what Bing Webmaster Tools' AI data can tell you. The free, first-party source this study is built on. The goal here isn't to judge one vendor; it's a reusable lens for the flood of "we did X and citations jumped" posts.
What did the case study report?
A large headline number, and a more complicated story underneath it. Over 19 weeks of Bing Webmaster Tools data. 11 weeks of baseline, the markup submitted in week 12, then 7 weeks after. The site reported the following, all attributed to waikay.io's own June-2026 write-up:
| Metric | Reported result |
|---|---|
| Weekly AI citations (avg) | 167 → 614 (~3.7×) |
| Citations per cited page (avg) | 3.65 → 8.55 |
| Peak weekly citations | 1,063 (week 19) |
| English-language raw volume | down ~8% |
| English citations by funnel stage | TOFU −80%, MOFU +44%, BOFU +406% |
| Share of total growth from non-English pages | ~75% (French + Spanish) |
The authors are commendably candid about the confounds: they note the non-English surge "likely" came from a deeper site re-crawl rather than the markup, and they list the limits themselves. A single domain, a short 7-week window, no controlled experiment. That candour is exactly what makes it a good teaching example.
The four questions to ask of any AI-visibility case study
Run every "we did X, citations rose" claim through these four. The waikay study passes some and openly fails others, which is the point.
- Is there a control, or just a before-and-after? A baseline isn't a control. If anything else changed in the window. A re-crawl, a freshness pass, a new backlink, an algorithm update. The before-and-after can't isolate the cause. Here the authors flag that a re-crawl plausibly drove ~75% of the lift, so the headline can't be cleanly credited to the markup.
- Who ran it, and do they sell what they measured? Independence matters. This is a vendor measuring the effect of its own product on its own site. Self-reported, single-domain. That doesn't make it false; it makes it a hypothesis from an interested party, which is weaker evidence than an independent or corroborated result.
- Did the valuable citations move, or just the raw count? Look at composition, not totals. English raw volume actually fell ~8%, while citations shifted down-funnel (commercial, bottom-of-funnel pages up sharply). Whether that's good depends on your goal, and it's invisible if you only read the 3.7× headline.
- Has it been replicated? One site for seven weeks is an anecdote, not a pattern. A finding earns weight when it reproduces across domains and time. Until then, treat it as a prompt to run your own test.
A 3.7× lift with no control, on one self-reported domain, where the vendor sells the thing being measured and a re-crawl could explain most of it, that's a hypothesis, not a proof. Read the limitations section as carefully as the headline.
What's the durable takeaway, separate from the headline number?
Two things survive the scrutiny. First, Bing Webmaster Tools is a real, free, first-party place to watch your own AI citations move: which is why a study could be built on it at all, and why you can run the same before-and-after on your own site instead of trusting anyone's. Second, citation totals and citation composition are different metrics. A page can deepen. Earning more citations per page, as this site did (3.65 → 8.55). Even as raw volume dips, which echoes the broader pattern that focused, well-covered pages tend to be cited more thoroughly. That's a citation-coverage story, and it's more useful than any single multiplier.
What doesn't survive is the causal leap. "We added markup and citations tripled" quietly becomes "the markup tripled citations", and the authors' own re-crawl caveat shows why that step is unearned. The same discipline applies to the entity-strength claims that fill this space: entity signals plausibly help, but proving it needs a control, not a coincidence.
How to test a claim like this on your own site
Don't take the case study's word; reproduce the method honestly:
- Establish a clear baseline in Bing Webmaster Tools and your server logs before you change anything.
- Change one thing at a time. If you ship markup and trigger a re-crawl and add content in the same week, you've rebuilt the confound you were trying to avoid.
- Watch composition, not just the total: track citations per page and which funnel stages move, the way the reporting stack frames it, so a flat or falling raw count doesn't hide a real shift.
- Give it time, and expect noise: short windows over-read random spikes; citation activity is volatile, and decays without freshness.
- Cross-check across engines. Bing is one surface; a lift there may or may not show in Google AI Mode, Perplexity, or ChatGPT.
Reading case studies well is the same skill as measuring your own AI visibility well: insist on a baseline, separate correlation from cause, and watch composition over time across engines. Turning scattered per-engine signals into one honest, cross-engine scoreboard, so you can test what actually moves your citations instead of trusting a headline. Is exactly what Buffy Intel is built to do. Questions: [email protected].