Original insights by Aja Frost
30-second rundown
Key learnings:
- HubSpot replaced traffic worship with commercial evidence: its AI Search work helped qualified leads from AI rise 1,850% while the brand became the most visible CRM in AI answers.
- Small technical and content experiments beat fashionable guesses: HubSpot tested each idea, abandoned weak tactics, and scaled only the changes that improved citations, visibility, accuracy, or leads.
- AI systems need facts they can find and read: pages tailored to a particular industry and use case, accessible pricing, faster-loading pages, and relevant mentions on other sites gave them better evidence.
This week: Create one page for a specific customer and need, such as “CRM for construction companies.” Answer the questions that buyer would ask, show how your product helps, and back it with a real customer example. Check that the key facts appear in the page’s HTML so AI crawlers can read them.
Coming 3 months: Build a small content-partnership programme with relevant industry publications, associations, or complementary businesses. Offer useful research, customer examples, or expert input that gives them a reason to mention your brand on their own pages. Track which mentions go live and whether your brand appears more often in relevant AI answers.
Aja Frost watched HubSpot’s famous SEO engine lose altitude, cough smoke, and attract a crowd of cheerful undertakers. Google changes and AI answers were stripping clicks from educational content. Frost’s response was not another strategy deck. She pulled two curious SEOs into Project Lighthouse, met with them daily, and shipped experiments every week.
By year-end, HubSpot was the most visible CRM in AI Search and qualified leads from AI had risen 1,850%. The team did not discover one silver bullet under a desert cactus. It built a testing machine, watched several ideas die noisily, and scaled the survivors.
1. The fashionable file went nowhere
Project Lighthouse first tested llms.txt, a plain-text map intended to help AI systems understand a website. HubSpot planted identifiable clues in the file, checked model answers, watched server logs, and even submitted it to Googlebot and Bingbot. Nothing useful happened. The bots left no fingerprints, and the grand new signpost stood in the sand pointing at itself.
The failure established the operating rule: test claims with your own data before scaling them. HubSpot paused the tactic instead of defending sunk effort, then moved its budget to experiments tied to citations, visibility, accuracy, referrals, or qualified leads.
That discipline also stopped the team from confusing movement with progress. AI share buttons increased citation rates by 29% across eight tested URLs, but visibility for related questions stayed flat. Seven pages improved on the intermediate metric, yet the result HubSpot cared about did not move.
2. Give the bots evidence they can actually read
HubSpot built 141 pages aimed at specific buyers, pairing an industry with a task its software could handle. Imagine a construction firm asking ChatGPT which CRM could help manage leads: a page about HubSpot for construction businesses gives it a more relevant source than a generic CRM page.
ChatGPT visited the pages roughly 15,000 times, but visits alone did little: early on, the pages appeared as citations in only about 16% of the answers HubSpot tracked. Later, HubSpot reported that 92% of the pages had been cited at least once and its AI-search visibility had risen 49%.
The pages became useful sources for specific buying questions; getting crawled was only the first step.
A different problem surfaced when HubSpot checked AI answers about its prices: they were often wrong. Google could read its JavaScript-rendered pricing pages, but AI bots could not reliably access those prices and drew on outdated third-party pages instead.
HubSpot published straightforward blog posts explaining each product’s pricing. Over two months, answers became more accurate for five of six products.
Their Sales Hub product was the exception: its answers grew less accurate until HubSpot also corrected old Sales Hub prices on outside websites. Your own page is evidence, but it is not the only witness in the room.
3. Speed and outside corroboration moved the needle
HubSpot tested pre-rendering, which served finished HTML to crawlers instead of making them assemble a heavy page. Load time fell 6.4 times to roughly one-tenth of a second. AI-bot crawls jumped 1,600%, citations rose nearly 40%, and AI referral traffic increased 6%. That gap is the whole crooked carnival: citations can surge while visits move only a little.
A fast, readable page got HubSpot into the evidence locker; mentions elsewhere gave LLMs more places to encounter the brand. HubSpot shifted its outreach toward getting its brand mentioned in relevant third-party content. It also invested in its subreddit. Community membership grew 61.7% year over year, HubSpot mentions across Reddit multiplied sevenfold, and citations doubled.
This is not permission to carpet-bomb forums with fake praise. HubSpot worked with its community team, earned co-moderator status, created useful programming, and recruited consistently helpful members as champions.
HubSpot’s playbook is less glamorous than the vendors selling instant domination, which is precisely why it matters. Pick a commercial question, form a measurable hypothesis, make the smallest credible change, and wait long enough to see crawling, citations, visibility, accuracy, referrals, and leads separately. The winners get scaled. The losers get buried before they eat the budget.
