SISTRIX published a test of which on-page GEO tactics go with AI citations on 22 October 2026, comparing 96,418 URLs across ChatGPT, AI Mode and AI Overviews. Pages meeting many of its eleven features were cited 2.1 times as often as barely optimised ones, and llms.txt showed no measurable effect, so content teams should fix page structure before adding files.
Key takeaways
- SISTRIX compared 48,527 pages cited by ChatGPT, AI Mode and AI Overviews with uncited pages from the same domains, 96,418 URLs from 17,380 domains in all.
- A table of contents scored 1.7 and Article schema 1.6, where 1.0 means no difference, yet only 14% of pages have a table of contents.
- Domains with an llms.txt file supplied 11.7% of cited URLs against 14.0% of all domains, so SISTRIX found no measurable effect.
- Only 10% of cited pages met many of the eleven features SISTRIX combined into an overall score.
- Advice and how-to content made up 48.7% of cited pages, and brand websites supplied 38.7%.
How SISTRIX ran the test
Johannes Beus, Julia Weissbach and the SISTRIX content team took 48,527 cited URLs from the SISTRIX AI Index for Germany, Austria and Switzerland in the second quarter of 2026, about 16,000 from each of ChatGPT, AI Mode and AI Overviews. They paired each one with a page from the same domain that wasn’t cited, then checked 43 features on every page.
Comparing pages within one domain takes domain authority out of the result. The SISTRIX write-up reports each feature as an odds ratio, so a score above 1.0 means cited pages had that feature more often and below 1.0 means less often.
SISTRIX also shows why small studies mislead. It ran the FAQ schema test on samples of three sizes and got three different answers, 23% less common on cited pages at 833 URLs, 31% more common at 4,600 and 19% more common at 96,418. The direction only settled from about 5,000 URLs, and the authors advise ignoring GEO studies smaller than that.
Which tactics showed an effect?
Most of the positive signals fall into two groups, pages that are easy to extract from and pages that show a real person stands behind them and keeps them current.
| Feature | SISTRIX score | What to do |
|---|---|---|
| Table of contents | 1.7 | Add one to long advice pages, since only 14% have one |
| Article schema | 1.6 | Mark up guides and advice pages |
| Five or more H2 headings | 1.6 | Break long pages into clear sections |
| FAQPage schema | 1.5 | Use it only for genuine questions |
| Table with a header row | 1.5 | Put comparisons and specs in a real table |
| Author box with photo | 1.4 | Name the author and show who they are |
| Question as an H2 | 1.4 | Phrase some headings as the buyer’s question |
| Visible last-updated date | 1.3 | Show it, and only change it when the content changes |
| Breadcrumb, HowTo or QAPage schema | 1.0 | Keep them for other reasons, expect nothing here |
| Recipe schema | 0.1 | No action, SISTRIX thinks AI prefers topic pages |
The platforms differ at the edges. ChatGPT scored 1.9 for question-style titles and 1.8 for tables with header rows. AI Mode scored 1.8 for a table of contents and for five or more H2s. AI Overviews scored 2.0 for Article schema and 0.5 for product pages.
Does llms.txt help AI citations?
Not in this data. Domains with the file supplied a smaller share of cited URLs than their share of the sample, and SISTRIX says llms.txt shouldn’t be a priority. Its comparison works at domain level, so it can’t rule out other differences between the sites that have the file and those that don’t.
Google says the same about its own products. Its AI features documentation tells site owners they “don’t need to create new machine readable files, AI text files, or markup” to appear in AI Overviews and AI Mode, and that there’s “no special schema.org structured data” to add. Article schema still scored well in the SISTRIX data, which probably says more about the care behind pages that carry it than about the markup itself.
Should shops change product pages?
SISTRIX’s split by site type points online shops somewhere else. For shops, Product schema scored 0.6 and a visible last-updated date 0.7, while brand websites scored 1.6 for both Article and FAQ schema and editorial sites scored 2.5 for a table of contents. SISTRIX’s advice to shops is an advice section with Article schema, rather than more work on product pages. Product schema scored 1.7 for SaaS providers, so the answer depends on what you sell.
Where content budgets should go
Our view at Manual Focus is that the SISTRIX study lets content teams drop the AI-only extras and put their GEO hours into page craft that Google Search rewards anyway, meaning clear sections, real tables, named authors and honest update dates on the advice pages buyers’ questions land on. These are the same habits we set out in our SEO cluster generator, and none of them needs a separate AI budget line.
The study also measures something narrower than visibility. A page can be cited without your brand being recommended, and AI answers increasingly cite pages Google doesn’t rank, which we covered in the citation moved off the page. Page structure decides whether an engine can lift your answer. Whether buyers then see your brand in those answers needs its own tracking.
The title result needs care. Pages with 2025 or 2026 in the title scored 2.7 and pages showing only an older year 0.7, but SISTRIX reads the year as a sign of fresh content and says it can’t confirm that changing the year alone helps. A team that edits the year on stale pages risks a misleading title and gains nothing the data supports.
Five fixes for your top pages
- Score your 20 most important advice pages. Check each for a table of contents, five or more H2s, a question heading, a table with a header row, an author box with a photo and a visible update date. Our site audit and refresh playbook covers how to run the sweep and ship fixes.
- Add Article schema to guides, and FAQPage only where the questions are real. SISTRIX’s 1.5 for FAQ schema held only for genuine questions, not token ones written for markup.
- Refresh before you re-date. Update the facts on pages with an old year in the title, then change the title and the visible date together.
- Take llms.txt off the roadmap. Leave an existing file in place, since other tools may read it, but Google says Search doesn’t need one, so don’t spend sprint time on it.
- Find the questions where rivals get cited. Bing’s Topics Portfolio report lists AI answer queries where other sites are cited and yours isn’t, which tells you which pages to fix first.
Where the study runs out
The study shows correlation, and SISTRIX says so. A table of contents may sit on pages that are better in ways the 43 features don’t capture. The data comes from Germany, Austria and Switzerland in one quarter, and UK results could differ. Perplexity, Claude and the Gemini app weren’t included, though SISTRIX plans a follow-up. Gemini 3.5 Flash classified domain and content types, with 84% agreement in a cross-check against Gemini 3.1 Pro, so the site-type splits carry more noise than the headline figures. The title findings were analysed after the fact and are labelled initial indications. The scores describe the second quarter of 2026, and the engines change often.
Further reading
- The citation moved off the page, on why Google rankings stopped predicting AI citations
- What Bing’s Topics Portfolio means for AI search content plans, for finding the queries where you aren’t cited
- SEO cluster generator, for building topic pages with GEO structure from the start
- Site audit and refresh, for running a page-by-page fix list through review
- AI Overviews hit 82% of tracked brand searches, on how often Google’s AI answers now sit on your brand terms
If you want help deciding which pages to rebuild first, see our services page.
Frequently asked questions
Does llms.txt help a website get cited by ChatGPT or Google AI Overviews?
Not measurably, according to a SISTRIX study published on 22 October 2026. Domains with an llms.txt file supplied 11.7% of cited URLs against 14.0% of all domains in the sample. SISTRIX says the file shouldn't be a priority, and Google's own documentation says sites don't need AI text files or special markup to appear in its AI features.
Which page features are linked to more AI citations?
In the SISTRIX study of 96,418 URLs, a table of contents scored 1.7, Article schema and five or more H2 headings 1.6, and FAQPage schema, tables with a header row and author boxes with photos between 1.4 and 1.5. A score of 1.0 means no difference. Pages that met many of eleven features were cited 2.1 times as often as barely optimised pages.
Does adding the current year to a page title get more AI citations?
Pages with 2025 or 2026 in the title scored 2.7 in the SISTRIX citation study, and pages showing only an older year scored 0.7. SISTRIX treats the year as a stand-in for fresh content and says it can't confirm that changing the year alone helps. Update the page first, then the title.
Should online shops optimise product pages for AI search?
SISTRIX's October 2026 study suggests not as a first step. For online shops, Product schema scored 0.6 and a visible last-updated date 0.7, and product pages scored 0.5 in Google AI Overviews. SISTRIX recommends shops build an advice section with Article schema, since advice and how-to content made up 48.7% of cited pages.
Sources
- Citation Worthiness, Which GEO Tactics Hold Up Across 96,418 URLs, SISTRIX
- AI features and your website, Google Search Central
From the podcast
James and Hannah go through the SISTRIX study published on 22 October 2026, which compared 96,418 pages to see which GEO tactics go with citations in ChatGPT, Google AI Mode and AI Overviews. Tables of contents, Article schema, real tables and author boxes went with more citations, llms.txt showed no measurable effect, and product pages did poorly. They cover how to read the title-year finding, what it means for online shops, why small GEO studies mislead, and three jobs to hand to AI.
Get AI on this
Paste these into whichever AI assistant you use. Swap the bracketed parts for your own details. The first needs an assistant that can open web pages, the second and third need file uploads.
1. Score your advice pages
I run marketing for [brand], [one line on what you sell] at [website], selling
mainly in [market].
A SISTRIX study of 96,418 URLs, published in October 2026, found these page
features went with more citations in ChatGPT, Google AI Mode and AI Overviews:
a table of contents, five or more H2 headings, at least one H2 phrased as a
question, a table with a header row, an author box with a photo, a visible
last-updated date, Article schema and FAQPage schema used for genuine
questions.
Here are our 20 most important advice and guide pages: [paste URLs].
Open each page and its source code and check each feature. Give me a table
with the columns: URL | Table of contents | 5+ H2s | Question H2 | Table with
header row | Author box with photo | Visible update date | Article schema |
FAQPage schema | Features missing (count).
Then rank the pages by features missing, most first, and for the top five list
the specific fixes. If you couldn't open a page or read its schema, say so in
that row rather than guessing.2. Match rival citations to your pages
I run marketing for [brand] at [website].
Attached is an export of Top Opportunity Queries from the Topics Portfolio
report in Bing Webmaster Tools. These are AI answer queries where other sites
are cited and ours isn't. Also attached is a list of our advice and guide
pages with their titles: [attach or paste].
For each query, find the page on our site that should answer it. Give me a
table with the columns: Query | Topic | Our best matching page (or "none") |
How well it answers the query (fully, partly, not at all) | What the page is
missing.
Group the queries with no matching page into suggested new pages, each with a
working title and the queries it would cover. Don't invent search volumes or
citation counts that aren't in the export.3. Plan an advice section for a shop
I run marketing for [brand], an online shop selling [product categories] in
[market].
A SISTRIX study published in October 2026 found that advice and how-to pages
made up 48.7% of pages cited by AI search, while product pages were cited
less often. I want to plan an advice section.
Attached are our product categories and the questions customers ask us most,
from [help-desk tags, reviews or site search, as available].
Propose 10 to 15 guide pages. For each give: a working title phrased as the
customer's question, the products it would link to, an outline of five or
more H2 sections, one comparison table it should include, and who on our team
is best placed to be named as the author.
Base every topic on the questions in the attachments, and say which
attachment each one came from. Flag any topic where you're unsure customers
actually ask it.And one for you
Take llms.txt off the roadmap. Leave an existing file in place, but put the sprint time into the pages above.
Read the episode transcript
James: If you've been sold a checklist for getting cited by AI, this week SISTRIX tested a lot of those tactics against almost a hundred thousand web pages, and some showed no effect at all. I'm James Vickers, this is Manual Focus, and Hannah has been through the study with me.
Hannah: Hi James. It's a good one to have read, because the method is better than most of what gets published on this.
James: Start there. What did they actually do?
Hannah: SISTRIX, the search data company, published it on the twenty-second of October. They took forty-eight thousand five hundred pages that ChatGPT, Google's AI Mode or AI Overviews had cited, from Germany, Austria and Switzerland in the second quarter. Then they paired each one with a page from the same website that wasn't cited, and checked forty-three features on every page.
James: Why does the pairing matter?
Hannah: Because it takes the size of the website out of it. A big publisher gets cited more whatever it does. Comparing two pages from the same site shows you what the page itself is doing.
James: So what came out on top?
Hannah: A table of contents scored one point seven, where one point oh means no difference. Article schema, the code that tells a search engine a page is an article, scored one point six. FAQ schema and tables with a proper header row were one point five. An author box with a photo was one point four.
James: And the pages that did most of it?
Hannah: Pages meeting many of eleven features were cited about twice as often as barely optimised pages. Two point one times. But only ten per cent of cited pages manage that.
James: Now the one everyone's asking about. llms.txt, the file some sites add to guide AI crawlers.
Hannah: No measurable effect. Domains with the file supplied eleven point seven per cent of cited pages, against fourteen per cent of all the domains in the sample. SISTRIX says it shouldn't be a priority. And Google's own Search Central documentation says you don't need AI text files or special markup to appear in its AI features.
James: So sites with the file come out slightly behind.
Hannah: Slightly, and SISTRIX is careful with it. It's a comparison between domains, so other differences between those sites could be in there. But nothing in the data says to put it on the roadmap.
James: There's a title finding that I suspect will be misused.
Hannah: There is. Pages with twenty twenty-five or twenty twenty-six in the title scored two point seven. Pages with only an older year scored nought point seven. SISTRIX reads that as a sign the content is fresh, and says it can't confirm that changing the year alone does anything.
James: So if your team edits the year on a stale page, you get a misleading title and nothing the data backs up. Refresh the page, then change the date.
Hannah: That's how I'd read it.
James: What about online shops? A lot of our listeners run ecommerce.
Hannah: Product pages did badly. For shops, Product schema scored nought point six. In Google's AI Overviews, product pages scored nought point five. Advice and how-to content made up nearly half of all cited pages, forty-eight point seven per cent. So SISTRIX's advice to shops is to build an advice section with Article schema, rather than keep working on product pages.
James: That fits something we wrote back in June, in the piece on the citation moving off the page. The overlap between Google's top links and the sources AI answers cite had dropped from around seventy per cent to under twenty. This study goes some way to explaining why. The engines want a different kind of page.
Hannah: And it measures citations, not recommendations. A page can be cited without the engine recommending your brand. So structure decides whether it can lift your answer, and you still need to track whether buyers see you.
James: Our view is that this lets content teams stop paying for AI-only extras and put the hours into page craft Google Search rewards anyway. Clear sections, real tables, named authors, honest update dates. None of that needs its own budget line.
Hannah: And SISTRIX has one more warning that's worth hearing. They ran the FAQ schema test on three sample sizes. At eight hundred and thirty-three pages it looked negative. At four thousand six hundred it was positive. The direction only settled from about five thousand pages.
James: So the next GEO study with three hundred pages in it goes in the bin.
Hannah: They'd put it more politely, but yes.
James: Right, so what do we hand over? Get AI on this. First, score your twenty most important advice pages against the study's list.
Hannah: Give your AI the URL list and the features to check, a table of contents, five or more section headings, a question as a heading, a table with a header row, an author box with a photo and a visible update date. It hands you a grid of which pages have what, and the fixes ranked by how many features each page is missing.
James: Second, find the questions where rivals get cited and you don't.
Hannah: Export the Top Opportunity Queries from Bing's Topics Portfolio report, which we covered a couple of weeks ago, and give it your page list. It matches each query to the page that should answer it, or flags that you haven't got one.
James: And for the shops, get AI on drafting the advice section. Give it your product range and your customer questions, and it hands you a list of guide pages and an outline for each. Whether they're worth writing stays with you. The prompts are in the show notes.
Hannah: And what we don't know. It's a correlation study, and SISTRIX says so. The data is from Germany, Austria and Switzerland in one quarter, so UK results could differ. Perplexity, Claude and the Gemini app weren't included, though a follow-up is planned. And the title findings are labelled initial indications.
James: So the takeaway. Make your advice pages easy to lift and clearly owned by a real person, and take llms.txt off the list. The full article, with the table of every tactic and its score, is on manual-focus dot co dot uk. If this was useful, follow the show in Apple Podcasts, Spotify or wherever you listen. And if you'd like help working out which pages to rebuild first, book a free working session with us at manual-focus dot co dot uk. Thanks, Hannah.
Hannah: Thanks, James. Go and check your tables of contents.