What AI Reads Before It Recommends: Inside 7,534 Perplexity Citations
AITwo Perplexity model variants were asked to recommend the five best tools in each of 380 B2B software categories. Their number-one picks matched in 79% of categories, and their top-five lists overlapped by 78% on average. That consistency is the interesting part: AI software recommendations are not noisy coin flips. They are a stable funnel — and a stable funnel can be engineered.
A dataset released today by Trellner Research (Manufactured sources behind AI recommendations, CC BY 4.0, collected 2 September 2026) makes the funnel measurable. This article walks through the raw numbers, adds an independent classification of the 2,055 cited domains, and points out a gap that should worry European software vendors in particular.
The experiment
The methodology is pleasantly verifiable. Trellner put 380 buyer-intent software categories — "project management software", "password managers", "CRM software" — to two web-grounded Perplexity models (perplexity/sonar and perplexity/sonar-pro) via OpenRouter. One prompt per category per model, 760 calls, zero errors. Each answer returned a ranked top five of vendors plus the grounding URLs the model reported retrieving. Every cited domain was then matched against the Tranco top-1M list and the Wayback Machine.
That leaves 3,800 recommendation slots, 7,534 citation rows, and roughly 9.9 grounding URLs read per answer. All numbers below come from that corpus; the domain-level classification is my own analysis of the same CSV.
AI reads the long tail — and sometimes nothing at all
If AI answer engines only trusted the established web, every citation would land on a top-10,000 domain. They do not:
| Metric | Value |
|---|---|
| Citations to domains not in the Tranco top 1M | 23.4% |
| Cited domains not in the top 1M at all | 751 of 2,055 (36.5%) |
| Median Tranco rank of ranked citations | ~71,600 |
| Share of citations going to top-10 domains | 17.3% |
Read that fourth row again. Over 82% of everything Perplexity reads before recommending software comes from outside the web's ten most-visited sites. The grounding layer is a long tail — and the long tail is where manufacturing costs are lowest. Of the unranked domains, 98.1% do exist in the Wayback Machine, so these are not dead links; they are live, purpose-built content properties.
Three kinds of source, one dominant pattern
Classifying all 2,055 cited domains by hand-tuned heuristics (multi-category footprint plus listicle URL patterns) yields three groups:
| Class | Count | Examples |
|---|---|---|
| Aggregator / media | 25 | g2.com, capterra.com, reddit.com, forbes.com |
| Multi-category content operators | 29 | guideflow.com, wifitalents.com, worldmetrics.org, gitnux.org |
| Vendor or niche content site | 2,001 | everything else, cited once or twice |
The 29 operators are where the study's title gets its meaning. wifitalents.com was cited 71 times across 27 unrelated categories — project management one day, applicant tracking the next. worldmetrics.org (60 citations, 22 categories) and gitnux.org (50, 23) run the same playbook: thousands of statistics-heavy listicles, generated at scale, structured so a retrieval pipeline can cite them cleanly.
The most instructive case is guideflow.com. It is a real vendor — an interactive demo platform — yet it appears in the grounding set of 96 of 380 categories, from video conferencing to A/B testing. Somewhere between those two extremes sits a whole industry: companies that treat "being the page the AI reads" as a distribution channel and build content operations around it. The practitioners call this generative engine optimisation (GEO); the dataset shows it working at industrial scale.
Stability makes it a market
Here is the part with real consequences. The two Perplexity variants are different models with different retrieval behaviour — yet their top-five sets overlap 78% (Jaccard) on average and their first pick is identical in 301 of 380 categories. Meanwhile, no vendor monopoly exists on the answer slots: the ten most-recommended vendors (microsoft.com, aws.amazon.com, google.com, adobe.com…) hold just 7.4% of the 3,800 slots. The funnel is concentrated at the source level — a few dozen content operators shape thousands of answers — but wide open at the vendor level.
Combine those two facts and you get the political economy of AI search in 2026: stable answers, fragmented winners, and a small set of gatekeepers deciding which grounding pages get read. If your category's grounding pages are owned by an operator, every model in the family inherits that answer. Fix the grounding layer, and you move every model at once.
The European blind spot
GraphWiz readers will care about this one. Among all 2,055 cited domains across 380 categories of B2B software, exactly one has a DACH-country domain: tanso.de, a Munich-based carbon accounting platform. One in 2,055 is 0.05%.
European B2B software vendors are, at best, nearly invisible in the AI recommendation layer — even as European buyers move their research workflows onto answer engines. For a market that has spent a decade building digital-sovereignty strategy, this is an uncomfortable inversion: sovereignty over the infrastructure means little if the recommendation funnel is written by content operators in another jurisdiction. European public-sector digital-sovereignty initiatives are steps on the infrastructure side; the grounding side is unwritten.
What to actually do
If you sell software, the actionable checklist from this data is short:
- Audit your grounding pages, not your rankings. Ask an answer engine in your category for its five picks and inspect the URLs it cites. If none of them are yours, that — not your SERP position — is your distribution problem.
- Publish primary evidence. The domains being cited are statistics pages and structured listicles. Original benchmarks, pricing tables, and comparison pages with clean HTML headings are what retrieval pipelines can quote; PDFs and marketing prose are not.
- Watch for multi-category operators in your niche. If a site with 20 unrelated categories ranks between you and your buyer, your category has been farmed.
- For buyers: click through. When an answer engine recommends five tools grounded on
wifitalents.com-class pages, the confidence of the answer is cosmetic. The grounding layer is visible — read it.
The full dataset, including per-category answers, all 7,534 citations, and evidence snapshots of the content operators, is available under CC BY 4.0 at trellner.com. The classification script and category-overlap analysis used in this article are reproducible from the CSV alone — an afternoon of pandas, and you can watch the machine read.
Shout-out: Chemistry Learning on Android — closed alpha, but already incredible
Away from the recommendation funnels: our chemistry-learning companion app is entering a closed alpha on Android, and the early builds are already a joy. If you want a first look, three steps:
- A — Join the Google Group: groups.google.com/g/graphwiz-testers
- B — Opt in on the web: play.google.com/apps/testing/ai.chemistry_learning_org
- C — Download on Android: Play Store — Chemistry Learning
Thank you to everyone testing — feedback from the first cohort shapes what ships.