👋 Hi, it’s Kyle and welcome to Growth Unhinged, my weekly newsletter exploring the hidden playbooks behind the fastest-growing startups.

HubSpot went viral last year, but for the wrong reason. It was reported that blog traffic fell sharply, driven by SEO changes at Google. Aja Frost, former marketing leader at HubSpot and now Director of Growth Marketing at Mercury, takes us behind the scenes into what happened and how HubSpot rebuilt their growth engine around AI search.

ICYMI: 2026 State of GTM survey | Growth Unhinged LIVE | Member Library

For over a decade, HubSpot was the industry standard for SEO. We ranked for hundreds of thousands of keywords and drove many millions of visits to our website every month.

Then, in late 2024, Google changed its algorithm — and HubSpot’s blog traffic fell off a cliff.

You might already know this, because the charts showing our decline went viral on LinkedIn, X, and industry sites.

What you probably didn’t know: Some of this decline was intentional. Alongside many legitimate, highly relevant keywords (like “sales compensation planning” and “content marketing templates”), we’d ranked for ones like “best OOO messages” and “office April Fool’s pranks.” These topics generated traffic but not conversions, and so for the last few years, we’d virtually stopped touching them.

We’d also shifted more and more of our resourcing from educational content — which we knew AI Overviews and ChatGPT were quickly commodifying — to “influence” content on YouTube, podcasts, and newsletters like The Hustle. This content was higher-funnel, so it converted more slowly, but it was also far more unique and built deep brand loyalty.

This didn’t make the traffic decline (or the surrounding conversation) easy to stomach.

Yet as I often tell my team, every headwind is a tailwind if you turn around.

AI was siphoning away our traffic, but it was also creating a brand-new channel that no one understood how to win yet. I decided to fast-track our AI search strategy. My goal: Recover the demand we’d lost and coin the next “inbound marketing” playbook for current and future customers.

I pulled two of our most AI-curious SEOs off their normal responsibilities and spun up an experimental pod called Project Lighthouse. We met daily and shipped AI search experiments weekly.

By the end of the year, HubSpot was the #1 most visible CRM in AI search and qualified leads from AI were up 1,850%. Here's how we got there.

At the time, everyone was hyping up llms.txt: a plain-text file that gives AI models an easy-to-read map of your website (like robots.txt, but specifically for AI).

We created one with a few “easter eggs” so we’d know if a bot had visited and ingested it. We also watched our server logs. Nothing happened: The models didn’t know our easter eggs when we asked, and our logs didn’t show any visits.

Next, we submitted the file directly to Bingbot and Googlebot to see if traditional crawlers would take the bait. Still nothing.

It was an inauspicious start, but it taught us an important lesson: Run your own tests. There’s so much noise in AEO — it’s critical to validate a play with your own data before scaling it. We paused llms.txt and agreed to revisit it only if new evidence came out. (Despite Google adding it to documentation and Chrome Lighthouse audits, a recent Ahrefs analysis showed 97% of llms.txt files get zero requests.)

Given how contextual AI answers are, we hypothesized we could drive visibility by publishing industry- and use-case-specific content: “CRM for [use case]” and “CMS for [industry.]”

We needed a lot of these pages, so we used AI to generate them: 141 pages, each targeting a specific industry × use-case combination and built on HubSpot's library of case studies.

The early batches were full of hallucinations, and the language was stilted and full of jargon. We were worried that even if the content could influence AI, it’d turn off any humans who found it — or worse, Googlebot. Our team kept a running QA doc, logging every fix and iterating the prompt to improve both accuracy and readability. Finally, we got the test batch to a good place.

We published the pages, and waited… and waited. AI bots were crawling the content (ChatGPT’s bot alone racked up 15K crawls in a few weeks), but citations hovered around 16%.

Then citations began to tick up. Which taught us the second lesson: Crawls happen first, then citations, then visibility. Ultimately, 92% were cited, increasing visibility by 49%. We expanded to additional industries and use cases and rolled out the same play for our German, Spanish, French, and Japanese sites.

Soon after, one of my team members spotted a problem: LLMs were giving people the wrong pricing. (A quick analysis of our sales call transcripts showed it was happening fairly often, too.)

We immediately knew the cause. HubSpot’s pricing pages are rendered with Javascript. Another team owns the pricing page, so we’d never been able to rebuild it. The architecture hadn’t been too problematic, since Googlebot can crawl JS. But AI bots can’t (yet), so they were grabbing our pricing from outdated third parties.

Knowing from our server logs that bots crawled our blog at a stunning clip, I suggested writing a series of straightforward blog posts. Each post would be structured for bots and describe a product’s pricing in-depth. (As a bonus, we had a lot more room than a pricing page.) It was far from scintillating content, but if it worked, we’d correct the bots’ HubSpot pricing knowledge.

And it did work. Over the next two months, accuracy significantly improved for five of our six products. (The one exception was Sales Hub, for which accuracy actually dropped. We closed the gap by correcting the outdated Sales Hub pricing on third-party sites, a time-intensive but reliable strategy.)

These past few experiments helped us arrive at a helpful guiding philosophy: Some content was meant for humans and bots. Some content was just for humans. Some content was just for bots.

Related: What AI agents really think about your pricing

We doubled down on the last category with our next launch. When we’d kicked off the pod, we’d intentionally focused on bottom-funnel content.

But in a brainstorming session, we came up with a way to make TOFU potentially work for us. What if we built a glossary for LLMs — each term getting a definition, example, and (this was key) one or two sentences explaining how it related to a HubSpot product? When an LLM was generating a response, it’d pull the meaning and the HubSpot mention. Or so we hoped.

To test this theory, we launched 50 glossary pages at hubspot.com/glossary, covering relevant terms like "audience segmentation" and "zero-click search." These pages were designed explicitly for bots: server-side rendered, HTML-first, with clear language and contextual HubSpot references. But like our industry- and use case-pages, we put effort into their design and content, so any humans who did find them would have a good experience, too.

The glossary lifted visibility 35% on awareness-stage questions and 26% on consideration/decision questions. Over the course of the pilot, HubSpot's overall citation share climbed from 1.97% to 3.2%. We published glossaries in five more languages and kept adding more terms.

Related: What Clay, Glean, Rippling and Lovable are doing to influence LLMs

At this time, we’d started seeing “summarize with AI” show up on competitor comparison pages. This tactic interested us for a few reasons — we hypothesized it could help train LLMs on an opinion about our brand (reinforcing the value props we were seeding across our own website and third parties) and potentially increase citations to the specific pages being summarized.

However, it wasn’t risk-free. Some SEOs had tried to hack AI responses by downvoting, at scale, all the negative responses about their brand and upvoting positive responses about their competitors, and this play seemed a little close for comfort.

To mitigate risk, we tested on lead magnet pages that were far from our core product pages. We also kept the pre-set prompt neutral: “Summarize the content at [URL].”

Citation rate increased 29%, with 7 of 8 URLs improving. However, visibility for related questions was flat. At the end of the day, we care most about driving visibility — so we decided to shelve this experiment to focus on more impactful ones.

We strongly suspected we could increase citations and visibility by serving pages to bots more quickly. To test this, we divided URLs into control and variant groups (controlling for traffic and related variables) and rolled out Botify SpeedWorkers (i.e. pre-rendering) to the variants.

This decreased load speed by 6.4X to roughly one-tenth of a second — and crawls exploded. We saw a 1600% increase from AI bots and, as a nice bonus, a 30% increase from traditional search crawlers. Citations increased nearly 40% and AI referral traffic increased 6%. (A good reminder that citation and traffic trends are not proportionate.) Based on this success, we expanded SpeedWorkers globally and across subdomains.

When the experimental pod started shipping on-site experiments, our link-building team had quickly pivoted into “mention-building.” Rather than commissioning third-party sites on direct sales for HubSpot, we paid them a flat fee to mention us. Now, mention-building is pretty common practice, but at the time, it was the least understood aspect of AEO. The team had to build our program and strategy from the ground up. After a few months of testing, they’d figured out:

At the same time, we were trying to crack the Reddit code.

It’s no secret that Reddit is one of the top-cited domains in AI search. The platform is a rich, constantly regenerating source of product reviews and has partnerships with both OpenAI and Google (although it might not renew the Google one…)

But Reddit is also famously hard for marketers to infiltrate. The community is self-moderated and allergic to “spam.” We did not want to piss Reddit off and spawn a ton of negative HubSpot comments.

Luckily, a subreddit already existed for HubSpot, and our mention-building team noticed something interesting: Whenever activity spiked in /HubSpot, positive mentions of HubSpot increased across Reddit. Our advocates were remembering we existed and they liked us, which spurred them to talk about us more outside the HubSpot forum.

With this principle in mind, we partnered with our Community team to make /HubSpot as engaging as possible. We successfully applied for co-moderator status; created a content calendar with product feedback prompts, practical how-to content, “HubSpot wins” posts, and more; and used AMAs and event programming to give people reasons to keep coming back. When we identified Redditors who consistently wrote high-quality, helpful comments, we invited them to become a “HubSpot champion.”

The community grew 61.7% year-over-year, HubSpot mentions across Reddit 7Xed, and citations doubled.

HubSpot knew traffic declines were coming for the rest of the internet, too. In part thanks to the great results we were seeing from Project Lighthouse, we acquired Xfunnel — the software my team had been using to guide our strategy and measure results.

HubSpot AEO launched in April, and its beta users saw impressive results. Our pivot worked.