What Sentry's jump from 15-20% to nearly 70% of signups from AI search means for growth engineers
Sentry's AI-search surge shows how to measure and earn citations.

Sentry’s reported jump from 15–20% to nearly 70% of signups attributed to AI search points to a change in what growth infrastructure includes. Awareness spending and content that machines can read now support the same path to discovery. For growth engineers, the opportunity is to measure that path, make the site citable, and test which awareness investments strengthen it.
Why Sentry’s AI-search number is a credible data point, not a marketing claim
Sentry’s figure deserves scrutiny because it is a named, first-party claim from a growth function that publishes its methodology. Matt Henderson’s April 2026 account of working without advertising cookies lets readers see how he chose to measure results. A signup attributed to AI search is still an attribution result, and its meaning depends on how the source was recorded.
Most B2B SaaS growth stacks were not instrumenting this channel a year ago, so many teams lack an equivalent baseline. A high reported share is useful as a reason to investigate. If Sentry’s figure holds up, what did the company do to get there, and what would it take to verify or replicate the result on another stack?
The structural shift that makes awareness spend and AI citation the same budget line
Henderson describes two returns from the same developer video. A person can watch it and remember the product, while an LLM can cite it when a prospect asks which monitoring tool handles a particular error. The source pool overlaps. Content made to build recognition can also feed model training and retrieval, giving awareness spending a role beyond influencing the person who first encounters it.
YouTube has overtaken Reddit as the most frequently cited social platform in AI-generated answers. That shift makes developer video relevant to citation as well as reach, but the broader implication is organizational. Awareness, SEO, and community increasingly contribute to one discoverability function. Treating them as separate budgets can hide the shared return, especially when each team measures only its own immediate traffic.
Meanwhile, fewer search referrals and more zero-click results weaken the old demand-capture model. AI Mode sessions overwhelmingly end without a visit to any site, so the site increasingly needs to supply an answer worth citing. But what happens when everyone floods the same sources with generated content? Henderson’s answer is that authenticity becomes more valuable as competent, generic material becomes abundant. Real people, deep technical knowledge, and an emotional connection help awareness work in ways that a markup change alone cannot.
What Sentry actually did — the four operational moves behind the number
Sentry allocated roughly 70% of its growth budget to awareness, including influencers, podcasts, out-of-home, Reddit, YouTube, and third-party newsletters. That is a different emphasis from the demand-capture spending common in comparable B2B companies. Multi-year sponsorships involving the Golden State Warriors, Valkyries, and Chase Center, plus sustained out-of-home spending, show the scale of its investment in recognition. Not every awareness placement supplies a directly citable source, even when it supports the same discovery path.
The second move was buying the Syntax podcast, a source developers trust that reaches them through audio, YouTube, and transcripts. The third was removing the traditional CMS in favor of a code-and-repo workflow for the marketing site. The fourth was building content for agents through llms.txt, structured data, readable pricing, and checkable proof. These moves address different parts of the path. Awareness helps people remember the company, the podcast gives them material they trust, and the changes to publishing and content help turn that material into answers they can use. The reported result does not isolate each move’s contribution, so the bundle is a case to examine.
Why Sentry removed its CMS and what that decision means for iteration speed
Henderson’s May 2026 migration post frames CMS removal as a way to shorten the distance between wanting a site change and publishing it. When content is part of citation infrastructure, a slow publishing process becomes a growth constraint. The decision is counterintuitive for a reason. A CMS usually lets non-technical editors make changes on their own, while a repository can shift the work to people with technical skills. Removing one bottleneck can create another.
For a team evaluating the same tradeoff, the question is whether its actual publishing cadence justifies the switch. A repo workflow needs technical editors or a process that keeps marketers involved without making engineers gatekeepers for every update. Otherwise, the apparent simplification may slow the people responsible for the content. Sentry’s choice is an argument for judging the workflow by how quickly accurate changes reach the site.
Sentry’s cookieless move offers a useful parallel. Both decisions were framed as forcing functions that required new capabilities to replace dependence on legacy tools. The relevant capability here is a repeatable publishing process that can keep machine-readable content current as training and retrieval windows change.
What content built for agents actually requires, beyond llms.txt
Henderson notes that bots now account for the majority of internet traffic. The response is not to stop writing for people, but to structure useful human content so machines can digest it. A file alone cannot accomplish that. The practical test is whether an agent can resolve a purchasing question from the site without requiring a person to click through and interpret missing details.
Sentry treats freshness as non-negotiable, using structured data and schema markup to surface an updated date on key pages. Readable pricing serves a different purpose. It lets an agent understand cost without a sales call, which is a prerequisite for the sign-up-in-chat possibility Henderson describes. Try-before-card access supplies a low-friction next step that bypasses a human-gated demo request.
Specific numbers, named outcomes, and verifiable claims are more useful than broad promises that a model can only paraphrase. Sentry’s content shift reflects this preference, moving away from top-of-funnel listicles toward how-tos, solutions pages, and in-depth tutorials that answer particular questions. Henderson’s October 2026 watermarking post adds a forward-looking concern as LLMs begin flagging AI-generated text. The durable response is proof grounded in human expertise.
How to instrument an AI-search attribution claim on your own stack
To test whether the same things happen elsewhere, a growth engineer needs three layers of measurement before using Sentry’s figure to make a business case. The first is self-reported signup source, with a “how did you hear about us?” field that explicitly includes AI assistants. This signal will underreport influence, but it captures sources that buyers remember well enough to name. It is an accessible starting point.
The second layer combines HTTP referrers from ChatGPT, Perplexity, Gemini, and similar services with server-log bot classification. Referrers identify direct visits, while classification separates agent crawls from human sessions. Those are different kinds of activity. Henderson flags that cookie-free journeys make last-touch ChatGPT visits likely to undercount the full influence path, because the upstream citation may not appear in the eventual visit. It is also still unclear why LLM-referred traffic converts differently from organic search.
The third layer measures visibility through systematic prompt queries across ChatGPT, Claude, Gemini, and Perplexity. Logged over time, these probes track citation velocity, platform distribution, and presence across clusters of adjacent questions. Because these probes measure inclusion, the dashboard needs distinct measures for AI impressions, LLM citations, share of AI answers, cross-model consistency, LLM-to-signup conversion, and LLM-to-customer conversion. Each needs an owner so visibility and revenue do not become interchangeable claims.
Internal data shows that LLM traffic converts at a much higher rate than non-brand SEO traffic. Conversion quality needs measuring even when AI referrals account for a small share of visits. It does not explain the difference by itself. A credible plan keeps the unresolved questions visible, including what self-report misses, what referrers omit, and how much observed citation presence translates into customers. None of these layers alone establishes the complete path from exposure to signup.
Why awareness spend without citation infrastructure leaks value, and citation infrastructure without awareness doesn’t get built
When buyers shop in AI Mode, brands can be filtered before the person sees the options. Recognition helps a company get through that filter, while structured, up-to-date content gives the model something it can cite. Those are separate jobs within one journey. Awareness without usable evidence can produce a name-drop rather than a linked recommendation, while an impeccably structured site does not create recognition by itself.
If the only expected return is an immediate click, maintaining agent-readable proof can look like secondary site work. Once the goal includes being recommended, that work has a direct purpose. The Syntax acquisition and developer video sponsorships supply material that can enter training and citation pools for future model versions.
Aggarwal et al.’s academic GEO research found that credible quotations, verifiable statistics, and explicit citations can raise source visibility in AI answers. The overlap with good human communication is useful. That suggests teams can write naturally for people while working to get cited by a particular model. Still, a model update can reshuffle citation patterns, leaving brands exposed if they have overcommitted to one platform-specific tactic. Spending on recognition and evidence people can trust is more likely to keep paying off, though neither ensures the same place in every answer.
Where a growth engineer should start — a prioritized sequence, not a full rebuild
The practical sequence starts with measurement before any content or budget change. A team should establish self-reported source, referrer and agent traffic analysis, and share-of-voice probes to capture its baseline. That gives later decisions a point of comparison. It also separates two problems that need different responses, namely not being named by models and being named without converting the resulting interest. Without that distinction, a budget reallocation is largely a guess.
The next step is a site audit for agent legibility, covering structured data, freshness signals, readable pricing, checkable claims, and llms.txt. These are low-cost changes that can ship in a sprint and help both people and machines understand the product. Content strategy comes next. The site needs enough how-tos, solutions pages, and in-depth tutorials to answer specific purchasing and technical questions. This is not a declaration that top-of-funnel content is dead, but a shift toward useful answers where broad search traffic may never arrive.
With that foundation, a team can choose one awareness channel where authenticity compounds in the developer community. A sponsored YouTube series, a podcast partnership, or a sustained Reddit or forum presence can be treated as long-tail citation infrastructure. The CMS decision belongs on a slower track. So does a wholesale budget reallocation, because both need separate business cases supported by the new instrumentation before proceeding beyond the first sprint’s scope.
Maintaining the work then becomes a question of cadence. AI content pipelines that have several stages and human review can cut the time for updates from days to minutes while keeping approval gates for brand claims and crawler policy. The autonomy ladder provides a boundary, with agents observing, analysing, drafting for review, and executing only allowlisted narrow tasks before stopping at consequential publishing decisions. Automation must preserve the ability to say no. With that boundary in place, each refresh whose results the team measures can add evidence people and models can use to the recognition the team is building, helping its next awareness investment do more than the last.


