Our AI Citation Experiment: Why 25% Was Not a Gain

We reviewed two DataEase AI scans for the same question and found a tempting story: our reported mention rate had gone from 20% to 25%. The underlying answers did not support calling that an improvement. Both scans had one engine naming us.

This is our first-party experiment baseline for the AI citation-tracking page, not a customer case study or a claim of uplift. We are documenting the observation, the correction prepared in this revision, and the evidence required before reporting a result.

What did we measure before changing the page?

We inspected the saved question "ai citation tracking for saas" on October 20 and October 27, 2026. The first scan returned 5 engines; the second returned 4. DataEase AI appeared in Gemini on both dates and in none of the other returned answers.

The October 20 sample included ChatGPT, Claude, Gemini, Perplexity, and Grok. The October 27 result for this question included ChatGPT, Claude, Gemini, and Perplexity. We do not infer why Grok was absent from the later result.

These are saved samples returned by the DataEase AI MCP, inspected on October 29. The dated engine counts and citation evidence are available for inspection. They are not all conversations on those engines and do not establish a market-wide visibility rate.

Why was the change from 20% to 25% misleading?

The numerator remained 1 mentioning engine while the denominator fell from 5 returned engines to 4. Restricting both dates to the 4 shared engines produces 25% on each date. The aggregate increased by 5 percentage points, but the matched comparison was unchanged.

This is a small example of why we do not want a dashboard percentage to become a growth story on its own. The missing engine changed the arithmetic. There was no new mentioning engine in the comparable sample.

The detailed wording also mattered. The later Gemini answer grouped our name alongside another tool. A positive name match did not establish that the answer described DataEase AI precisely. Description accuracy needs its own review.

Did a citation mean the answer recommended us?

No. Perplexity cited our AI visibility tools roundup on both dates but did not name DataEase AI in either answer. That is 1 observed owned-source citation in each Perplexity sample, distinct from a brand mention or a recommendation of our product.

The cited source was our tools comparison, not the citation-tracking landing page. That distinction changes the next task. We should inspect what each page contributes to the answer rather than treating any link to our domain as proof that the product page worked.

We preserve three fields in this review: the engine named us, the engine cited our domain, and the description was accurate. Combining them too early would hide the work that remains.

What exact correction are we making?

In this revision, we add 1 dated evidence section to the citation-tracking page, separate mentions from source citations, and link to the monitoring method. We also replace its free-analyzer wording with Brand Readiness language so the entry point matches the tool's actual scope.

The original gap was not a missing URL. The landing page already existed, yet a saved recommendation still asked us to create it. The better brief is to improve the existing explanation and show what a real sampled result can and cannot tell a founder.

The hypothesis is modest: a concrete example and accurate scope should help a reader understand the workflow. Whether either engine cites the updated page is a separate outcome. We do not predict a score lift.

Can we attribute the October scan results to a website refresh?

No. The page displayed an October 21 update label, but that is not a verified deployment timestamp. The two scans also returned different engine sets. Without a confirmed intervention time and repeated comparable observations, these 2 samples cannot establish a content-driven effect.

The repository history and page metadata are evidence about source revisions, not proof of when an engine fetched the deployed content. We therefore use the October samples as context and a baseline audit, not as a before-and-after success story.

This matters for the correction in this article too. Its measurement clock begins when the revised content is deployed and verified, not when we prepare the draft.

How will we decide whether the correction helped?

After deployment, record the live publication time and keep the question wording fixed. Compare the same 4 shared engines across at least 2 follow-up scans. Review exact cited URLs and description accuracy, and report missing runs separately. Keep search traffic and lead actions outside the mention-rate calculation.

• Verify the updated page is publicly accessible and record the deployment time.
• Retain the October 27 answers as historical context; take a fresh baseline if deployment is delayed.
• Inspect the first two scheduled post-deployment scans using the same prompt and engine cohort.
• Record whether the landing page, the tools roundup, or another URL was cited.
• Report unchanged, improved, worsened, or insufficient evidence. Do not count a missing run as a negative answer.

Other site changes and engine variability remain possible explanations even if a later sample improves. This is an observational experiment, not a randomized test.

What can other founders take from this baseline?

Keep 3 distinctions visible: readiness versus observed answers, mentions versus citations, and a changed percentage versus a changed result. Save the underlying evidence before publishing the story. A smaller claim that survives inspection is more useful than an apparent gain caused by a missing denominator.

Our LLM monitoring walkthrough shows how to build the baseline. The measurement guide explains comparison windows, and the DataEase AI platform connects findings with reviewable work through Branding and supporting AI Agents.

As of October 29, 2026, this article reports the baseline and prepared correction. No post-deployment uplift is claimed.

For a worked action sequence, follow the AI citation playbook: inspect a question, review the evidence, and choose one improvement.