What 48,527 AI-cited pages have in common, and what that means for a CPG website

Most advice about getting cited by AI assistants comes from small tests, single case studies or lists of the most-cited sites, which mostly tell you that big brands are big. On October 22, SISTRIX published something more useful: a comparison of 96,418 URLs that holds the website constant and changes only the page.

The design is the reason to pay attention. SISTRIX took 48,527 pages from its AI Index for the German-speaking market that had been cited by ChatGPT, Google's AI Mode or AI Overviews in the second quarter of 2026, roughly 16,000 from each. Each cited page was paired with a page from the same domain that wasn't cited. Because both pages share a brand, a domain and a reputation, what's left to explain the difference is the page itself. The study covered 17,380 domains and checked 43 features on every page.

It's still a correlation study, and SISTRIX says so plainly. A page with a table of contents might be cited more because of the table of contents, or because the kind of team that adds one also writes better pages. But it's the best large-scale look we have at which of the usual tactics line up with citations, and a few results cut against what CPG brands are often told.

How to read the numbers

Every result in the study is a factor. A factor of 1.7 for the table of contents means that of two pages on the same domain, the one with a table of contents was cited 1.7 times as often. A factor of 1.0 means no difference, and anything below 1.0 means fewer citations. SISTRIX notes that these are technically odds ratios, which at a citation rate around 1.4% behave almost exactly like a straight ratio.

Show the data

Feature	Citation factor
Table of contents	1.7x
Article schema	1.6x
FAQPage schema	1.5x
Table with a header row	1.5x
Author box with photo	1.4x
Visible last-updated date	1.3x
TL;DR at the beginning	1.3x
Breadcrumb, HowTo, QAPage schema	1x
Product schema	0.8x
Recipe schema	0.1x

Source: SISTRIX, Citation Worthiness study, October 2026 (DACH region, Q2 2026)

The single highest number in the study was 2.7, for pages with 2025 or 2026 in the title. SISTRIX treats that one carefully: the title and H1 features were added after the original analysis, and the year most likely stands in for content that is new or kept up to date. Whether changing the year alone does anything can't be told from the data. We've left it off the chart for that reason.

What goes with more citations

SISTRIX boils its results down to two ideas, and they're a good summary. Make the page easy for a machine to pull apart, and show that a person stands behind it and keeps it current.

On the structural side, a visible table of contents with jump links had the strongest association, at 1.7, and it was also rare: only 14% of pages had one. Article schema came in at 1.6, and in most content systems it's a single setting. FAQPage schema came in at 1.5, across all three platforms and every domain type except encyclopedias. That last point matters because Google limited FAQ rich results years ago, and plenty of teams took the markup off. SISTRIX's condition is that the questions are real, not filler. Smaller signals included tables with a header row (1.5), at least one question used as an H2 heading (1.4) and a TL;DR at the top (1.3).

On the trust side, an author box with a photo came in at 1.4, and a visible last-updated date, at least ten external links and at least three types of structured data each came in at 1.3. Only 5% of pages showed a last-updated date at all.

None of these is large on its own. Combined, they add up: pages meeting many of eleven features SISTRIX grouped into a score were cited 2.1 times as often as barely optimized pages, and mid-range pages 1.6 times as often. Only 10% of cited pages already met many of them.

What didn't hold up

Three results are worth a CPG founder's attention, because they cut against common advice.

llms.txt showed no measurable advantage. Because the file applies to a whole domain, SISTRIX compared domains rather than pages here. Domains with an llms.txt made up 11.7% of cited URLs, while they were 14.0% of all domains in the data. If anything, they were slightly underrepresented. SISTRIX doesn't claim the file hurts, since those domains differ in other ways, only that it shouldn't be a priority. We publish one on this site and still agree; our own tip on llms.txt already called it optional, and this is the first large dataset to back that up.

Not all schema is equal. Breadcrumb, HowTo and QAPage markup showed no difference at 1.0, even though breadcrumbs were on 43% of pages. Recipe schema came in at 0.1 across every industry with enough data. SISTRIX's hypothesis, which it says hasn't been validated, is that AI systems prefer to cite a topic page rather than a single recipe, and its suggestion for food publishers is to build topic and overview pages with Article schema as citation targets. That's directly relevant to food brands with recipe sections.

Readability scores and word counts are symptoms. A good Flesch reading score came in at 1.03, which wasn't significant. Pages over 1,000 words came in at 1.5, but SISTRIX points out that longer pages tend to carry more headings, tables and lists. Cover the topic properly; don't write to a word count.

The part that matters most for brands that sell products

The platform and site-type breakdowns are where the study gets interesting for anyone with a shop.

Show the data

Feature and segment	Citation factor
Article or FAQ schema, brand websites	1.6x
Last-updated date, editorial sites	1.9x
Product schema, all pages	0.8x
Product schema, online shops	0.6x
Last-updated date, online shops	0.7x
Product schema, AI Overviews	0.5x

Source: SISTRIX, Citation Worthiness study, October 2026

Product schema came in at 0.8 overall, 0.6 on online shops and 0.5 in AI Overviews. SISTRIX's reading is that AI Overviews mostly cite informational content rather than pages built to sell. A visible last-updated date, which helped editorial sites (1.9), went with fewer citations on shops (0.7). Meanwhile, almost half of all cited pages (48.7%) were advice and how-to content, and the largest share (38.7%) came from brand websites.

Put those together and the advice for shops is unusually direct. SISTRIX writes: "For shops, an advice section with Article schema is therefore more promising than optimising product pages."

We'd add one caution before anyone deletes their Product markup: don't. SISTRIX says Product schema has purposes beyond AI citations and should stay, and our own schema tip still applies. Product markup and a Merchant Center feed are how your price, stock and GTIN reach shopping results and agents. What this study measures is something different: which page gets quoted as a source when someone asks a question. That page is rarely a product detail page.

What this looks like on a CPG site

A lot of CPG brand sites are a home page, a shop, an about page and maybe a blog that stopped in 2023. The study suggests the missing piece is a small set of pages that answer the questions shoppers bring to assistants, built with the structure that goes with citations.

Feature (factor)	What it looks like on a CPG site
Table of contents (1.7)	Jump links at the top of a guide like "How much protein is in a serving of our bars, and how that compares"
Article schema (1.6)	Article or BlogPosting markup on every guide page, with a real author and date
FAQPage schema (1.5)	Five to eight questions customers actually ask, answered on the page, marked up to match
Table with a header row (1.5)	A real HTML table comparing your flavors, sizes or ingredients, not an image
Author box with photo (1.4)	The founder or formulator, two lines on why they know this
Question as an H2 (1.4)	Headings like "Is it safe for kids with nut allergies?" instead of "Allergens"
TL;DR at the top (1.3)	A three-line answer before the long version

A protein snack brand might start with three guides: one on how its protein content compares per serving and per dollar, one on allergens and how it's made, and one on which product suits which use. A personal care brand might cover ingredients it leaves out and why, how to choose between its formulas, and how its certifications work (tip 21 covers how to show those). Each page gets a table of contents, a table, a real FAQ, a named author and a date.

Pew's 2025 browsing study gives a reason to care about being the source even when few people click. It found that Google users clicked a link inside an AI summary on just 1% of visits to pages with one. The answer itself does the persuading, and the brands quoted in it shape what it says. Being the cited source is how your facts, rather than a reseller's, end up in that answer.

Small studies, big swings

One finding is worth keeping for the next time a vendor sends you a GEO study. In SISTRIX's test runs, FAQ schema looked harmful at 833 URLs (23% less common on cited pages), strongly helpful at 4,600 (31% more common) and moderately helpful at the full 96,418 (19% more common). Only from around 5,000 URLs did the direction hold steady. SISTRIX's rule of thumb is not to trust a GEO study with fewer than 5,000 URLs, and it applies that to its own subgroup results too.

The study also has limits worth stating. It covers the German-speaking market, three platforms (no Perplexity, Claude or Gemini chat) and one quarter. Classification into domain and content types was done by a Gemini model, cross-checked with a larger Gemini model (84% agreement) and a manual sample of 100 pages. A follow-up is planned. US results could differ, though the underlying behavior of the systems is the same.

What we'd do this quarter

If we ran a CPG brand site with a few hours a week to spare, the order would be:

1. Turn on Article schema for every blog or guide page. It's usually one setting, and Google's Article structured data documentation lists the fields.
2. Add a table of contents and a TL;DR to the guide pages you already have. Both are template changes, done once.
3. Write three advice pages around the questions customers ask, each with a real FAQ (tip 7), a comparison table (tip 22), an author and a visible date (tip 23, tip 29).
4. Leave Product schema alone. It's doing a different job.
5. Skip the llms.txt debate. Publish one if you like; it takes fifteen minutes. Don't count it as progress.

Then check which pages assistants actually cite and for which questions. You can do that by hand, as in tip 12, or with a tracking tool that works at the product level, such as SearchPanel from XENA Intelligence, which asks assistants shopper-style questions and keeps the answers. The point is to see whether the advice pages start turning up as sources. If they do, you'll have built the thing our earlier piece called a reference book for robots, and the grocery site data suggests most of your competitors haven't.