Skip to content
PMD

Case studies

Case studies of the week: local AI Overviews, GEO tactics that get cited and AI citations that come back

Three dark cards with the figures 99%, 14% and 75%, from Semrush, SISTRIX and Ahrefs studies on AI Overviews and AI citations.

This week’s four studies agree on one thing: visibility in AI depends on what others say about you and on how your page is structured, not on holding one position or on a single isolated tactic. All four come from SEO tool vendors analysing their own data, so read them as indications, not laws.

What happens to the map pack when AI Overviews become almost universal?

In a Semrush study published on 17 November, Chris Shirlow ran 480 searches by hand: 12 US cities (from big metros to small towns), 10 types of local business (home services, legal, healthcare and restaurants), two ways of searching and two rounds between September and November 2026.

The headline result: in the first round 38% of searches showed an AI Overview (the AI-generated summary Google places above the results); by the last round, 99% did (79 of 80). On the 145 searches where the summary appeared above the map pack (the three-business map block), it left out at least one pack business 114 times (79%). More than a third of the businesses it recommended (373 of 1,023) were not on Google’s first page, and it cited sources such as Yelp, industry directories, Zocdoc, Facebook and Reddit.

What readers can apply: being in the map pack no longer guarantees the AI names you. Check which sources list your business (directories, reviews, profiles) and that their data matches your site. The author states no limitations; it is a one-person study with manual searches, US only.

Which GEO tactics are really associated with being cited?

SISTRIX analysed 96,418 URLs from 17,380 domains in Germany, Austria and Switzerland in the second quarter of 2026. It compared pages cited by ChatGPT, Google’s conversational AI search mode: you chain questions together and get composed answers with links to the sources. and AI-generated summaries Google shows above the results for some searches, with links to the sources it uses. with uncited pages from the same domain, across 43 characteristics.

What comes out: pages with a visible table of contents and jump links are cited 1.7 times as often, and only 14% have one. Article markup goes with 1.6 times as many citations and FAQPage with 1.5, even though Google dropped FAQ Google results with extra elements (ratings, prices, FAQs, images) that appear thanks to a page’s structured data.. Having 1,000 words or more goes with 1.5 times, but SISTRIX warns this mostly reflects better structure. And on Proposed file at a site’s root that lists its most useful pages in Markdown so language models can find them. It is a proposal, not an official standard.: 11.7% of cited URLs come from domains with that file, against 14.0% of all domains, so no measurable effect for now.

What readers can apply: start with the cheap, structural fixes (a table of contents with anchors, clear headings, Article and FAQPage) before files like llms.txt. These are correlations from one region, without Perplexity or Claude, as the study itself acknowledges.

Which kind of mention carries the most weight in AI answers?

Ahrefs matched 10,000 brands against 260 million unbranded prompts (study of 3 November, US) and used Measures whether two variables rise and fall together, from −1 to 1, without requiring a straight-line relationship. The closer to 1, the more they move together; it does not prove one causes the other. (a measure, from 0 to 1, of how much two figures tend to rise together) to see which external mentions go with more mentions in AI. Reddit mentions in ranking threads lead with 0.75, followed by domain and web page mentions (0.70), news (0.67), YouTube (0.64) and TikTok (0.60). A brand with 10 times more Reddit mentions has about 5.5 times more AI mentions; with web pages the effect flattens: 10 times more mentions means about 3.7 times. The median brand has just 8 Reddit threads and 29% have none.

What readers can apply: map where your brand is discussed outside your site (forums, media, video) and work where there is room. It is correlation, not causation, with data from Ahrefs’ own tool.

How long do AI citations last?

Another Ahrefs study, from 29 October, followed about 37,500 citations across five AI tools, asking the same questions every day for 20 days. By the next day at least half of the citations had disappeared: around 60% on AI Mode and ChatGPT, 50% on AI Overviews and 80% on Gemini and Copilot. But roughly 75% of those that disappeared came back within 20 days. Only 14–25% of one-day citations returned the next day, against 69–87% of those that had already lasted eight days or more.

What readers can apply: do not react to a one-day drop in citations. Measure over multi-day windows and with many questions, as we explain in how to measure AI visibility. The study’s own limitations (only the first disappearance was tracked, and the rankings analysis uses 230 questions) call for caution.

What we take away this week

  1. Look after third-party sources, not just your site: profiles, directories, reviews and mentions in forums and media.
  2. Structure every key page: a table of contents with anchors, question-style headings, Article and FAQPage.
  3. Prioritise what is measurable over what is fashionable: llms.txt showed no effect in the SISTRIX study.
  4. Measure over long windows: a citation lost today often returns, and a stable citation is the signal that matters.
  5. If you run a local business, check what AI Overviews say in your area even if you already rank in the map pack.

To apply this to your brand, our GEO service starts with a diagnosis of how AI cites you today. If you would rather talk first, get in touch. For the big picture, start with what GEO is.

Shall we talk about your project?

The first consultation is free. Tell us where you are and we will tell you where to start.