I thought the GEO tests from Hubspot were super interesting, especially with the additi
Key Takeaways
- HubSpot's "Project Lighthouse" AEO programme (case study published by Aja Frost, 16 September 2026) drove a 1,850% increase in qualified leads sourced from AI over 12 months.
- Pre-rendering pages for bots using Botify SpeedWorkers cut load time roughly 6.4x (to around a tenth of a second), producing a 1,600% increase in AI-bot crawl volume, a 30% increase in traditional search-crawler activity, and a near-40% increase in citations.
- An llms.txt file produced no measurable engagement from AI crawlers in HubSpot's testing.
- A 50-term glossary built specifically for LLM consumption lifted visibility by 35% on awareness-stage queries and 26% on consideration/decision-stage queries.
- 141 hyper-specific "industry plus use case" pages were cited by AI systems 92% of the time, lifting visibility by 49%.
- Zyppy's 2026 Google Ranking Factors Expert Survey (131 SEO professionals, 103 factors, AI-answers cut published by Cyrus Shepard, 16 September 2026) rated llms.txt files at just +0.05 impact, the lowest-scoring factor, while Brand/Entity in LLM Memory scored +2.08 and AI Crawl Access and Snippet Eligibility scored +2.20, the highest.
- Search Engine Journal's brand-protection framework (Olesia Korobka, 15 September 2026) proposes a four-layer defence sequence for how a brand is represented across ChatGPT, Gemini, Perplexity and Claude.
- Google's John Mueller confirmed (16 September 2026, via Search Engine Roundtable) that Search Console's AI Mode position-tracking methodology is still evolving and is not a fixed standard.
- Google's new Data Manager tool is showing an average 26% increase in incremental ROAS for offline/app conversions and an average 11% increase in Search conversions via Enhanced Conversions, per Google's own announcement (10 September 2026).
- Andy Crestodina's Orbit Media 2026 blogging survey found marketers combining six or more proven strategies are nearly three times more likely to report strong results, while AI usage alone (92.4% of respondents) correlated with nothing.
What actually works in AI search: HubSpot's year of experiments
Most AEO advice is still theoretical. What Aja Frost has published is not: a detailed, honest account of a year-long internal programme at HubSpot testing specific tactics against specific outcomes, with the failures included alongside the wins. That combination is rare enough to be worth taking seriously.
The standout result came from infrastructure, not content. HubSpot used Botify SpeedWorkers to pre-render pages for bots, converting dynamically rendered content into static HTML so crawlers did not need to execute JavaScript to see it. Tested properly, with control and variant groups rather than a blanket rollout, the effect was substantial: load time fell roughly 6.4x, AI-bot crawl volume rose 1,600%, traditional crawler activity rose 30%, and citations rose nearly 40%. For any site with meaningfully JavaScript-rendered sections, filter pages, dynamic pricing or availability, single-page-app style templates, this is worth auditing before anything else on this list. A crawler that cannot render a page quickly may simply not bother with it, no matter how good the content underneath is.
Content-side, a few tactics earned their place and a few did not. A 50-term glossary built specifically for how LLMs parse and retrieve definitions lifted visibility 35% on awareness-stage queries and 26% on consideration and decision-stage queries. 141 narrow "industry plus use case" pages, built with AI-assisted generation, were cited 92% of the time and lifted visibility 49%. AI share buttons increased citations 29% but produced flat visibility gains and were quietly shelved. And despite a year of industry hype, an llms.txt file produced no measurable engagement from either AI models or crawlers at all, a result that lines up with Zyppy's survey data below.
Net effect over twelve months: qualified leads sourced from AI up 1,850%, and HubSpot became the most visible CRM in AI search, a result significant enough that it led to acquiring Xfunnel and launching a dedicated "HubSpot AEO" product line in April 2026. The caveat worth stating plainly: these are the results for one large, well-resourced B2B SaaS site. The specific percentages will not transfer directly to a different industry or a smaller site, but the ordering of priorities, infrastructure first, then structured and extractable content, then deliberate off-site presence, is a genuinely useful template.
The AI-answers cut of the big ranking-factors survey
Cyrus Shepard has published the AI Overviews and AI Mode-specific cut of Zyppy's 2026 Google Ranking Factors Expert Survey, the same 131-expert, 103-factor dataset whose traditional organic-ranking cut circulated a fortnight earlier. Two results stand out for being non-obvious. First, llms.txt files scored just +0.05, near the very bottom of the list and effectively rated as ineffective by the surveyed experts, directly corroborating HubSpot's own experimental finding above. Second, Brand/Entity in LLM Memory scored +2.08, the third-highest factor overall, ahead of several more traditionally emphasised technical factors.
The practical read, in the surveyed experts' own words: "First, answer the query head-on and structure the answer so the system extracts it without effort. Second, get your brand into the LLM's memory." That second point is doing more work than it might first appear. It is not a technical fix so much as a sustained presence and consistency requirement, being mentioned accurately, repeatedly, and in the same terms across enough of the web that a model's internal representation of a brand becomes stable and correct.
Auditing brand representation across AI search
Search Engine Journal's brand-protection piece pairs naturally with the LLM-memory finding above. Olesia Korobka lays out a practical four-layer defence sequence for managing how a brand is actually represented across ChatGPT, Gemini, Perplexity and Claude: fix what is within direct control, correct what can be influenced but not directly edited, report outright violations and impersonation, and publish clearer first-party answers where a fix or correction is not possible. The audit itself is comprehensive, covering markets, languages and devices, search surfaces, the AI systems themselves, autocomplete, paid ads, and both direct and decision-stage queries.
Related, and worth flagging on its own: Google appears to be testing AI-generated descriptions inside local Google Business Profile knowledge panels, an "AI Overview" label that expands into a chat-style interface. One SEO who tested it on his own listing found the generated description mostly accurate but slightly dated, a discrepancy he attributed to his own ageing website content rather than the AI misrepresenting anything. It is a useful, low-stakes illustration of exactly the control problem the brand-protection framework above is built to address: the AI system is not inventing information, it is surfacing whatever it can find, correct or not.
Search Console's AI Mode reporting will keep changing
Asked directly whether individual citations within an AI Mode response are counted sequentially or as a single position, Google's John Mueller was candid that the underlying reporting methodology is itself still being worked out: "The goal is not a written-in-stone absolute truth for position counting, that's impossible, but rather to make something that's useful for site owners in understanding how their site is shown." Worth remembering the next time an AI Mode metric moves without an obvious cause: the measurement itself may be what changed, not the underlying visibility.
A real measurement upgrade for first-party data in Google Ads
Google's new Data Manager tool provides direct GA and DV360 integration for first-party data and is showing an average 26% increase in incremental ROAS for offline and app conversions, alongside an average 11% increase in Search conversions attributed to Enhanced Conversions. Its Data Manager API now follows the IAB Tech Lab's ECAPI standard. Alongside it, a new Data Strength Uplift metric quantifies how much of that lift is specifically attributable to first-party data infrastructure, averaging around 14% uplift with Google Tag Gateway and over 20% for Demand Gen campaigns.
Meridian, Google's free open-source Bayesian marketing-mix-modelling tool, also picked up three upgrades worth knowing about: AI-assisted data-quality auditing, brand-signal integration such as branded Google query volume, and its geo-experimentation feature, GeoX, is now generally available worldwide at no cost. For anyone not currently using a marketing-mix model, GeoX's global availability is probably the most immediately actionable item here.
The old-school tactic getting a second look: HTML sitemaps
Eli Schwartz makes a strong case for a proper HTML sitemap, not the XML kind. An XML sitemap is a machine-readable manifest with limited value for actual discovery. An HTML sitemap is a real page that a human or a crawler can navigate, organised by theme, keeping every page within roughly three clicks of the homepage and eliminating orphan pages entirely. He cites measurable crawl and indexation gains at SurveyMonkey and Scribd following launch, and points out that LinkedIn, TripAdvisor and Amazon all quietly run HTML sitemaps of their own, LinkedIn's is visible only to logged-out visitors, effectively reserving it for crawlers, and paginated by entity type.
There is a genuine AEO angle here too, one that connects several of this week's stories. AI crawlers and LLM retrieval systems behave like a third audience alongside humans and traditional search crawlers, one that fetches and extracts rather than browsing menus or crawling patiently. A flat, plain-language HTML sitemap gives that third audience exactly the same easy, legible map of a site's content that it gives humans and traditional crawlers. Given the emphasis this week's ranking-factors survey places on AI crawl access and extractable content structure, an HTML sitemap is a concrete, buildable step toward both.
More from that big blogging survey: effort still wins
Following up on Andy Crestodina's Orbit Media survey: the more useful story in the data is not decline, it is what actually separates marketers still getting strong results from everyone else. Original research increases the odds of strong results by roughly 50%, and yet fewer marketers are doing it. Influencer and expert collaboration is the single best predictor of success. Marketers combining six or more proven strategies are nearly three times more likely to report strong results, even as AI usage on its own, now used by 92.4% of respondents, correlates with nothing either way. The honest conclusion is that complacency and shortcuts are what is losing ground, not the underlying channel. The people still putting in real, structured effort are seeing it pay off.
Google's own data: ChatGPT shoppers still return to Google before buying
A Google-commissioned survey of over 12,000 people who use ChatGPT for shopping research found that 99% of them still consult Google Search to help finalise the purchase decision. This is worth reading two ways. It helps explain why OpenAI continues pushing on ChatGPT's own advertising, monetising the research phase since people are not yet transacting inside the chat itself. And it is a reminder that AI chat interfaces are not yet the full purchase journey. Trust and established habit are still pulling people back to search before they commit to a purchase, an advantage that may not hold indefinitely as AI-native shopping matures, but holds today.
Google tests ads that look just like normal AI Mode answers
Google is testing a new AI Mode ad format built around text-anchor sponsored links, styled to look almost identical to the ordinary text links that appear within an AI Mode response, labelled "Sponsored" above rather than set apart visually as a distinct unit. Early signal suggests a lower click-through rate than other AI Mode ad formats despite the more native appearance, a useful reminder that blending in is not automatically the same thing as performing better.
Chrome adds new ad-measurement metrics to CrUX
Chrome's DevTools team has added four new experimental ad-measurement metrics to the Chrome User Experience Report: Ad Count, Ad Density, and Ad Weight by network and by CPU. These give anyone running third-party ads real aggregated user data for valuing ad inventory or evaluating a publisher partnership, without relying on guesswork. Worth being precise about what this is not: Google's own post is explicit that these are not part of Core Web Vitals and carry no suggested targets or thresholds, so they should not be treated as a ranking signal.
Google Ads may be expanding message assets to the ad-group level
Google Ads already supports linking ads out to Messenger, WhatsApp, Zalo or SMS conversations at the account and campaign level. A recent report suggests this is being tested at the ad-group level too, offering finer-grained control over which ads carry a messaging option. Google's own documentation has not yet caught up to confirm the ad-group-level detail specifically, so it is worth treating as an early or unofficial rollout for now rather than a fully documented feature.