Here’s my round-up of the key audience growth stories that caught my attention this week.
Let’s leap in.
The Beers & SEO with Barry and Steve Podcast is back!
It’s another turbo-charged (well, beer-fuelled) edition of our podcast, as Barry and I cover some of the biggest stories in SEO. This time, we take a closer look at the new functionality in Google Search Console that lets publishers remove their content from AI Overviews, AI Mode and AI-powered Discover.
Prediction: the next massive Google update is around the corner
Lily Ray has her crystal ball out, although, in fairness, Google has written most of the prediction for her. Its documentation now says it is critical to manually fact-check and review all AI-generated content, and describes fake author profiles as a form of deception. Spam updates are speeding up too: there have been four so far in 2026, compared with just one in the whole of 2025, while the gaps between them have shrunk from 92 days to 55, then 37. And at Search Central Live in Barcelona, Gary Illyes said scaled content is becoming more of a problem than link spam.
Why it matters: If you publish AI-assisted content at scale, or your author pages are thin, audit them now. Lily’s advice is to remove templated pages that aren’t performing and make sure every byline belongs to a real person with genuine credentials. Essential reading.
Google Tests Paying Publishers for AI Answers via Search Console
Matt G. Southern reports on a pilot that Google has confirmed to Digiday. Publishers whose content ‘contributes significantly’ to answers in the Gemini app, AI Overviews and AI Mode can get paid, with monthly earnings shown in a new Search Console panel. Digiday says at least dozens of publishers have been approached, and they aren’t all news organisations. The panel doesn’t explain how the earnings figure is calculated, which one executive described as ‘quite black box’. Another source said the early returns are small compared with advertising revenue.
Why it matters: If an invitation arrives, read the terms before clicking accept. One executive warned that taking the money could weaken your position in any later negotiation, because Google could point to the pilot as evidence that payment has already been made.
Why we don’t do prompt tracking
Dan Petrovic’s argument is that prompts are endless variations on a relatively small number of core intents, so his team tracks those instead. Each topic, ‘running shoes’, for example, is sent to Google, OpenAI and Anthropic models using one fixed prompt, repeated up to 100 times because the answers vary from run to run. He also introduces ‘selection rate’, the AI equivalent of click-through rate: how often a model cites a page from the set it was given. In one test, Google was given seven pages and cited all seven. OpenAI was given 39 and cited two.
Why it matters: Last week, Peec AI told us to track more prompts for longer. Dan says don’t track prompts at all. They do agree on one thing: a single answer on a single day tells you very little.
Fan-Out Mapper: Explore AI Search Subqueries
Neat idea, this. Brandleap’s tool starts with Google Autocomplete suggestions and maps the subqueries that sit around a topic. A heatmap shows which of them your pages already cover, while a ‘Battle mode’ compares your coverage with a competitor’s. Credit to Brandleap for saying plainly that this is an approximation built from suggestions, not a window into Google’s actual query fan-out. You have to request access, and there’s no pricing on the page.
Why it matters: A pretty clever way of approaching topical authority.
Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion
One for the brave. Pengcheng Jiang and ten co-authors examine searches where the goal is to retrieve a good set of results, judged on diversity, coverage and how well the items complement one another, rather than simply finding the single best match. Their method, R4T, trains a lightweight retriever to perform the fan-out in a single pass. On fashion and music benchmarks, it improved retrieval quality and cut fan-out latency by an order of magnitude.
Why it matters: This is a research paper, not a description of how any live search product works. But it points towards an interesting direction: systems designed to assemble a varied set of sources. Being the page that adds something different to that set may matter more than being the tenth version of the same answer.
Re-imagining AI Visibility KPIs in a Query Fan-Out World
Wil Reynolds argues that appearing in an AI answer is worthless if that answer ultimately steers people away from you. His example is ChatGPT telling users to look beyond Airbnb and VRBO. Banana Republic, meanwhile, ranks on page one of Google for ‘ethical jeans’ but never appears in the AI answer, despite Wil saying he has checked 1,000 times. He proposes five measures, including traffic from places where people have actively chosen to share your work, mentions in trusted trade publications, and how often you appear when the same prompts are run 10 to 30 times a month.
Why it matters: For publishers, the lesson is that reputation is increasingly measured outside your own site. Who mentions you, and how they describe you, may shape whether AI recommends you or merely names you.
Q2 AI Intelligence Report 2026
Comscore’s latest quarterly report uses its own panel data to track how people use AI tools. It covers which assistants are gaining users, mobile usage growing faster than desktop, sponsored ads in ChatGPT, how people prompt chatbots and changes in AI citations.
Why it matters: Panel data records what people actually did, making it a useful check on the many AI visibility figures built from synthetic prompts.
Draft Study: Users Assigned AI Mode Clicked Out Less and Trusted It Less
Alessandro Benigni writes up a draft paper from researchers at the University of Pennsylvania and Northeastern University. They enrolled 1,444 people and tracked 1,100 of them for seven days using a browser extension. Participants were split between normal Google Search, Search with AI Overviews hidden, and AI Mode. The AI Mode group clicked through to websites less often. They also trusted Search less, found it less useful and satisfying, carried out fewer searches each day and were more likely to try a rival search engine. Hiding AI Overviews increased clicks, with no measurable reduction in how people rated the experience.
Why it matters: It’s a draft that hasn’t been peer reviewed, so the findings may change. If they hold up, it provides experimental evidence for what publishers have long suspected, with the awkward twist for Google that users may not prefer AI Mode either.
Google Search Central Live Deep Dive, Barcelona – Day 1 Recap
John Campbell at ROAST took notes in Barcelona so the rest of us didn’t have to. Gary Illyes said AI Mode queries have doubled every quarter since launch and are around three times the length of a traditional search query, while one in six uses voice or images. Cherry Prommawin and Gary also explained that Search, AI Overviews and AI Mode all use the same crawler, Googlebot, while Gemini has its own. There was also a warning that CDNs are blocking more bot traffic, and that new Cloudflare settings can accidentally block Googlebot. My favourite moment: when asked what to do about Googlebot crawling stale cached pages, Gary’s answer was: ‘Nothing, just chill.’
Why it matters: Check your CDN and bot protection settings. If they’re challenging Googlebot, you could be cutting yourself out of traditional Search, AI Overviews and AI Mode at the same time.
Google Shows How Long Crawling, Indexing and Recovery Can Take
Matt G. Southern picks out perhaps the most useful slide of the week, again via John Campbell’s notes. Gary Illyes gave typical timings for some of Google’s processes: around 20 hours to discover a new URL, roughly 1.5 hours to index a page end to end, one to three months for a site move, and three to six months to recover from a core update. The slowest cases can stretch to a year or more, or even to ‘never’. Matt points out the caveats: the slides haven’t been published, and there’s no sample size or definition of what Google means by ‘typical’.
Why it matters: Handy numbers for the next time someone senior asks when the traffic is coming back. Treat three to six months for recovery from a core update as a planning assumption, not a promise.
Google AI Overviews Jump From 26% to 80% of Branded Queries
This is something we covered in our latest podcast. Danny Goodwin reports DemandSphere data showing AI Overviews appearing on more than 80% of tracked branded queries by late September, up from around 26% at the start of the month. The peak was 90.48% on 27 September. Separately, Chris Long of Nectiv tested 100 enterprise brands and found an AI Overview for 93 of them. Google hasn’t announced a change, and neither dataset measured clicks, so the effect on click-through rates remains unknown.
Why it matters: Search for your own masthead and see what Google’s AI says about you and which sources it cites. Brand searches used to be the one results page you could rely on owning.
AI Search Should Complement SEO Strategy, Not Replace It
Kyle Sutton, Head of AI Discovery and SEO at The Washington Post, gives four reasons not to bin your SEO strategy. Search is still the second-largest source of publisher traffic behind internal traffic, according to Chartbeat, while AI platforms account for less than 1% of publisher pageviews. Similarweb found that 95% of ChatGPT users also use Google. And many of the tactics are largely the same, which is also what Google’s own guidance says.
Why it matters: It pairs neatly with Gary Illyes’s line from Barcelona. If The Washington Post is keeping its SEO fundamentals and adding AI measurement on top, that seems a sensible model for the rest of us.
Opinion: SEOs Should Not Be Responsible for the “Agentic Web”
Joe Hall has noticed that anything touching search visibility tends to end up on the SEO team’s desk sooner or later: page speed, JavaScript, UX, digital PR, structured data. He’d like the agentic web, where AI agents act on behalf of users rather than simply fetching information, to stay off the pile. Agents that complete transactions need authentication, payments, inventory checks and security. Those are jobs for product and engineering teams. His argument is that SEO should help agents discover and understand information, but shouldn’t own the infrastructure that allows them to operate the business.
Why it matters: Publishers will face the same question about who owns agent readiness. Decide that now, before it lands on whoever last mentioned AI in a meeting.
AI Visibility Data Paralysis: How to Get Unstuck
Mateusz Makosiewicz argues that ‘AI visibility’ is really several different problems. You may be missing from the conversation, losing the comparisons, being described using out-of-date information or blocking the crawlers. Each problem needs a different fix. He uses Ahrefs’s own figures to show how one overall score can hide important gaps: Ahrefs is mentioned in 96.2% of answers about the best SEO tools, 8.6% for local SEO and 0% for social media. His team also asked more than 20 publishers to correct outdated information about Ahrefs. A few agreed, and those corrections later appeared in AI answers. It’s a 34-minute read and something of a tour of Ahrefs products, so pack a lunch.
Why it matters: Split your tracking by topic before reporting a single visibility score to anyone. An average can look healthy while the subjects you most want to be known for are completely missing.
The LLM Positioning Lag: What It Is and How to Avoid It
Dana Nicole gives a name to something you may already have noticed: you change what your business does, but the chatbots carry on describing the old version. Duolingo added maths in 2022, music in 2023 and chess in 2025, yet ChatGPT still described it as a language-learning app. Hotjar came second for heatmapping software with no mention of Contentsquare, despite the two companies completing their merger by July 2025. The fixes are unglamorous: announce the change clearly, update old pages and keep your description consistent across LinkedIn, Wikipedia, directories and your own site. The examples are individual ChatGPT answers selected by the author rather than a study, and Semrush has a tool to sell you at each step.
Why it matters: Any publisher that has rebranded, launched a new product or changed its paywall should ask the major assistants to describe it. Years of old coverage can outweigh last month’s announcement.
Until next time,
Steve
Appreciate the share, Steven! Big fan of your work.