We measured our own AI citation share for 90 days. Here is what moved it.

Nearly every "AI visibility" vendor shows you a case study with a number going up and no way to check it. We wanted the opposite: a measurement we run on our own sites, publish in full, and let anyone reproduce with the same prompts. So over roughly 90 days we tracked how often ChatGPT, Claude, Perplexity and Gemini cite our domains when people ask the real buying questions in each category. Here is what we found — including the parts that didn't work.

What we measured, exactly

The metric is Share of Citation (SoC): across a fixed set of buyer-intent questions in a category, the percentage of AI answers that name a given domain as a source. We ran it across four engines — ChatGPT (OpenAI), Claude (Anthropic), Perplexity, and Gemini — on 180+ prompts per category, and re-ran the identical prompt set on a schedule so the trend line is honest. Same questions, same engines, every time.

The headline: 0% to 13.1%

Our clearest result is TaxStand, a DIY Texas property-tax tool. At the first baseline it was cited in 0% of AI answers about protesting property taxes — dead last among twelve tracked brands. Ninety days later it sat at 13.1% Share of Citation: second in its category, having passed a firm that has been in business since the 1970s.

It's not the only one. Measuring the tool that runs these audits against its own competitors, it holds the top Share of Citation in its category — ahead of larger, better-funded names. And the rest of the portfolio spans the honest range you'd expect: some sites still sit near 0% against category leaders in the 20%+ range. That's the real starting point for most sites, and it's exactly why measuring first matters.

The four levers that actually moved the number

Across every site that climbed, the same four things did the work — in this order:

1. We measured before we touched anything

Every site started with a baseline of the exact buyer questions and a score of who the engines cite today. That turns "improve our AI presence" into a specific list of gaps against named competitors. Without it you're guessing, and guessing is what most retainers quietly bill for.

2. We made the site legible to machines — and verified it

Crawlability, clean headings, real structured data, and answer-style content so an engine can read and quote the page. The unglamorous unlock was verifying indexation, not just submitting URLs: across the portfolio, roughly 61% of sitemap URLs were confirmed indexed by Google. Pages that aren't in the index can't be retrieved, and pages that can't be retrieved can't be cited — no amount of on-page work rescues them.

3. We earned citations with disclosed, genuinely useful answers

We published direct, honest answers to the questions the engines were actually asking, on the sources they pull from — no overlays, no fake reviews, no spun filler. Corroborated, quotable facts are what the grounding step lifts. (The mechanics are in the companion piece, How AI decides what to cite.)

4. We re-measured in the open

The same prompt set, re-run on a schedule, so we could tell a real climb from noise and catch regressions. 0% to 13.1% is only meaningful because it's the same 180+ questions each time.

What didn't move it

• Publishing more, faster. Volume for its own sake added weak candidates, not citations. Answering one real question well beat ten pages that circled it.
• Schema in isolation. Structured data helped machines parse pages that were already good; on thin pages it changed nothing.
• Chasing an engine we weren't in. When a site was invisible everywhere, per-engine tactics were a distraction — the fix was retrieval and corroboration, which lifted all four at once.

The instrument

Every number here comes from CitedAudit, the same tool we point at client sites: multi-engine probing, Share of Citation, competitor benchmarking, and verified indexation. We built it because we needed an honest meter for our own portfolio before we'd sell anyone else a reading. The full, self-updating results live on its public proof lab, and the TaxStand climb has its own detailed case study.

If there's one takeaway: your AI citation share is a number you can measure this week, and the first honest reading is almost always lower than you'd like. That's fine — it's the same place TaxStand started.