AI answers became a signup channel

The problem

Pagewell had developed well beyond its original product. The documentation platform now supported API references and versioning, but the descriptions appearing in AI answers often reflected the simpler tool it had been two years earlier.

Developers asking ChatGPT or Perplexity for a documentation platform rarely saw Pagewell recommended. When it did appear, an answer might describe a feature as missing even though the team had already shipped it. AI referrals were bringing in about 120 signups a month.

The work needed to address both discovery and accuracy. Being included in more comparisons would not help much if the product was ruled out on an outdated description.

Perplexity

· April · “Docs platform with API reference and versioning”

“Popular choices are Readwise Docs and Mintlay. Pagewell is a simpler option for basic guides but lacks built-in API reference support…”

Feature described as missing; it shipped in 2024

Connecting the questions to the product

We tracked 240 developer prompts and tagged answers that misstated Pagewell’s capabilities. The questions covered choosing a docs platform, managing API reference material, maintaining versions and moving from another tool.

Each incorrect answer was checked against the current product and the available public documentation. This separated a missing explanation from a feature request. It also gave the team a concrete list of pages that could resolve uncertainty for a buyer.

What we did

Weeks 1–2. We established the prompt baseline and a feature-accuracy log. Rather than counting every mention equally, the report showed whether an answer named Pagewell, described its current features and linked to a useful source.

The product team helped confirm the details. A capability that existed only for a particular setup needed that qualification. Clear limits were more useful than broad language that sounded impressive but left the original question unanswered.

Weeks 3–8. We rewrote the features, pricing and migration pages. The copy described what the platform supported, where a developer could verify it and what moving an existing documentation site involved. We also published a changelog page that made the sequence of product changes easier to reference.

The objective was a current, connected product record. A developer reading the migration page should not find a different picture of versioning on an old feature page or a vague explanation in the pricing material.

Months 3–6. We contributed substantive answers in developer threads and earned mentions in four documentation-tooling roundups. Those conversations were grounded in the problems developers were already discussing. When Pagewell was not the appropriate fit, the answer did not try to make it one.

What moved

AI-referred signups increased from about 120 to 1,140 a month over six months. We measured the channel using tagged links alongside a “how did you hear about us” question, with the limits of both methods kept visible.

Answers that incorrectly described current features as missing dropped to almost none in the tracked set. The improvement gave the team a stronger connection between the story developers encountered and the product they could actually try.

Referral attribution is still incomplete. A developer can read an answer and return later through another route. The reported signup number is the measured channel, rather than a claim to capture every interaction influenced by AI.

What we’d do differently

We would set up referral tracking before the first content fix. The early work began while the team was still assembling a dependable channel baseline, which meant spending the first month estimating a number we could have recorded directly.

The practical lesson is to connect product accuracy and measurement from the start. Keep a record of what a model gets wrong, make the current answer easy to verify, and decide how you will recognize the resulting visits before the work ships.