Where do different LLMs send users for software searches?

Key numbers

What we found, in six numbers

How we ran it

Method and limitations

We wanted to know which kinds of sites AI answers cite when someone is close to choosing software, and whether that changes by search type and by AI engine.

• Searches. 148 real searches with measurable US search volume (Google Ads data via DataForSEO, at least 20 searches a month), frozen before any AI answers were pulled. 5 software categories (observability, CI/CD, project management, email marketing, help desk) and four kinds of search: “best X tools” (48), “X alternatives” (24), “A vs B” (25) and “A B integration” (50). One search (“bamboo alternatives”) was dropped because the engines answered about the plant and 3D printers, leaving 147.
• Engines and runs. Google AI Overviews (from the Google results page, desktop), ChatGPT and Gemini (the answer as the consumer app shows it, web search not forced) and Perplexity (through its API, model “sonar”, web search on). Each search ran 3 times on each engine: 1,764 answers. Data pulled between Oct 31 and Nov 3, 2026.
• Claude. The same 147 searches ran once each through Claude (claude-sonnet-5-5, via the API with web search on): 147 answers and 964 citations. Like Perplexity, this is the model called through the API; the claude.ai app may cite differently. With one run, Claude has no stability number. Every “all engines” figure on this page is the original four engines; Claude is always shown on its own.
• Prompt and citations. The prompt was the exact search, nothing added. A citation is a unique page cited in one answer (tracking parameters removed). 28 ChatGPT, 16 Gemini and 7 Claude answers cited nothing; 2 searches showed no AI Overview.
• Site types. Assigned by rules (known platforms, review sites, media, software vendors) plus a manual review of every other domain cited twice or more. Vendor pages are split by URL into blog & content, product pages and docs; that split is approximate. Domains cited only once (1.9% of the four engines’ citations) are left unclassified.
• Named product and mentions. A citation is “named” when it’s on a domain owned by a product in the search. For mentions, we matched about 80 product names against the answer text and checked whether each named product’s own site was cited in the same answer; that analysis is directional.
• Limitations. A short snapshot, and answers vary from run to run. US English only. Perplexity and Claude results come from their APIs, which can differ from the apps. A share of citations is not a share of clicks or traffic, and the data shows what gets cited, not why. Labels are rule-based plus manual review and can contain small errors.

Search type 1 · “Best X tools”

Software companies write most of the shortlist

On “best X tools” searches, 64% of citations were software companies’ own pages, and 90% of answers cited at least one company’s blog post. Third-party sites still show up: 85% of answers cited a review, independent or media site. ChatGPT went to product pages (66% of its citations); Gemini and Perplexity cited roundup posts (65% and 62%), and Perplexity leaned on media (13%, led by Forbes, PCMag, TechRadar).

Search type 2 · “[Product] alternatives”

Other companies write the answer about you

On alternatives searches, 70% of citations went to other software companies and 2% to the product being searched. 99% of answers cited another company; only 14% cited the product at all. On ChatGPT, other companies were 84% of citations. Google AI Overviews cited a Reddit thread in 73% of these answers, more than on any other search type. The product was named in 96% of these answers’ text.

Search type 3 · “A vs B”

A third company often frames the matchup

37% of comparison citations came from a software company that was neither A nor B, and 91% of answers (259 of 285) cited at least one such page. The engines split: on ChatGPT, 63% of citations came from A and B themselves; on Gemini, 8%.

Companies cited most often on other companies’ “A vs B” searches

Search type 4 · “A B integration”

For integration searches, docs are the answer

66% of integration citations came from the named products’ own sites and 39% were docs and help-center pages (64% on ChatGPT). 89% of answers cited a docs page from one of the two products. Google AI Overviews also cited a YouTube video in 88% of integration answers.

Five engines

Each AI engine reads a different internet

Google AI Overviews cited YouTube in 89% of answers and Reddit in 45%. ChatGPT and Perplexity cited neither and leaned on software companies’ own pages (84% and 82% of citations, against 39% for Google). Claude, via the API, cited independent comparison sites more than any engine: 19% of its citations, and at least one in 65% of its answers. 65% of the pages Claude cited had “2026” in the title, against 22% for ChatGPT; that’s a year-stamped title, not proof of a publish date.

What it means: there isn’t one “AI search.” Videos and Reddit threads matter for Google, your own site for ChatGPT and Perplexity, and independent comparison sites for Gemini and Claude. And because answers change between identical runs, a single check can mislead.

The Zapier pattern

zapier.com is the second most-cited website, and almost none of it is in dev tools

zapier.com was cited 466 times, second only to youtube.com, in 388 of 1,764 answers (22%), and in 31% of alternatives and vs answers. Most citations were blog posts titled almost exactly like the search (about a third were on a URL containing the exact search) and pages for connecting two specific apps. Of the 466, 258 were email marketing, 148 project management, 53 help desk, 5 CI/CD and 2 observability. In dev-tool categories other vendors fill that role: Octopus Deploy, Semaphore, Harness in CI/CD; SigNoz, Dash0, New Relic in observability.

Medium and dev.to

Medium and dev.to barely register

Medium was 0.6% of the four engines’ citations and dev.to 0.2%; ChatGPT and Perplexity cited no Medium posts at all. Of the 52 Medium and dev.to posts the four original engines cited, 44 were written by individuals and 8 by company accounts. Buyers may trust founders and engineers more than brands, but AI answers mostly cite company-owned sites (69% of the four engines’ citations).

What to do with it

What software marketing teams should do

1. Search your own name the way buyers do. Run “[your product] alternatives” and “[your product] vs [a similar tool]” in each engine your buyers use, a few times each, and note whose pages get cited.
2. Publish the comparison and alternatives pages you’d want a buyer to read. Specific about where you win, honest about where you don’t.
3. Make your homepage and pricing page say plainly what you do. ChatGPT cites product pages directly on “best” searches.
4. Get onto the roundups and review sites cited for your category. Most engines name products from someone else’s page.
5. Treat integration docs as a buyer-facing page. Public, both product names, plain-text setup steps.
6. If your buyers use Google, plan for YouTube and Reddit. Google AI Overviews cited them in most answers; the other engines didn’t.
7. Measure AI visibility over repeated runs. Answers changed between identical runs, especially on Gemini.

Data explorer

Explore the data

Download

Download the data

The site-type share tables (every engine × search type) and the top 60 domains, as CSV.