👋 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