Short answer:
- The four major AI assistants mostly agree on which brands to recommend, but they almost never cite the same sources.
- ChatGPT cites product pages for more than half of its sources, while Claude, Gemini and Perplexity rely mainly on third-party "best X" and comparison articles.
- ChatGPT adds the current year to most of the searches it runs behind the scenes.
- Asking the same question twice gives noticeably different answers.
If you want AI assistants to recommend your company, you're not optimizing for one search engine. You're optimizing for four retrieval systems with different habits.
How we ran the study
In November 2026 we wrote 70 questions a real buyer of marketing software would ask before choosing a tool. They covered:
- category questions ("What is the best AI SEO tool?")
- comparisons and alternatives ("Profound alternatives for AI search visibility")
- problem-first questions ("How much does SEO cost per month for a small business?")
- industry questions ("Best marketing software for healthcare practices")
We ran them, with web search on, through:
- ChatGPT: the live chatgpt.com experience, run twice for 40 of the questions to measure consistency
- Claude (Sonnet 5.5)
- Gemini (3.5 Flash)
- Perplexity (Sonar Pro)
That produced 230 answers. For each answer we recorded the brands mentioned, every cited URL and, where available, the background searches the model ran ("fan-out queries"). Data was collected through DataForSEO's AI Optimization API.
Finding 1: The engines agree on brands, not on sources
For the same question, two engines named the same brands about 41% of the time on average.
But the overlap in cited domains between any two engines was just 6%:
| Engine pair | Overlap in cited domains (same question) |
|---|---|
| ChatGPT ↔ Claude | 3% |
| ChatGPT ↔ Perplexity | 5% |
| ChatGPT ↔ Gemini | 6% |
| Claude ↔ Perplexity | 7% |
| Claude ↔ Gemini | 11% |
| Gemini ↔ Perplexity | 4% |
This lines up with other research putting cross-engine overlap in the single digits to low teens. Practical meaning: "rank in AI" isn't one goal. A page that gets you cited in Perplexity may do nothing for ChatGPT. Track each engine separately.
Finding 2: ChatGPT cites product pages; the others cite roundups
We grouped every cited URL by page type:
| Page type | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Vendor homepage / product page | 52% | 5% | 7% | 9% |
| "Best X" lists, alternatives, comparisons, reviews | 22% | 64% | 56% | 52% |
| Blogs and guides | 7% | 21% | 16% | 20% |
| Official platform docs (Google, OpenAI, etc.) | 6% | <1% | <1% | <1% |
When ChatGPT recommends a tool, it usually links straight to the vendor's own page. Claude, Gemini and Perplexity mostly link to the third-party article where they found the recommendation.
Practical meaning:
- For ChatGPT, your own product and pricing pages need to state clearly who you're for, what you cost and how you compare.
- For the other three, you need to be in the roundups and comparison articles they trust. Publishing more on your own blog won't get you there.
Finding 3: ChatGPT adds the current year to most of its searches
When ChatGPT searches the web to answer a question, it often runs several background queries. 76% of the ChatGPT background queries we captured (152 of 201) included "2026". Some examples:
- "best ai cmo tools for startups 2026 comparison autonomous marketing strategy execution pricing"
- "jasper ai alternatives 2026 ai writing tools comparison copy.ai writesonic anyword rytr"
- "best ai search visibility tools for startups 2026 ai search optimization monitoring platforms pricing features"
Claude added the year less often (34 of 67) and Gemini rarely (18 of 63).
Studies have found the median page AI cites is over a year old, so this is not a reason to rewrite "2025" as "2026" in your titles and call it done. It does mean that pages that are actually current, and say so, line up with how ChatGPT searches. That covers pricing, feature comparisons and "best of" lists especially.
Finding 4: Ask the same question twice, get a different answer
We ran 40 questions through ChatGPT twice, with the same wording on the same day.
- The two answers shared 62% of the brands they mentioned.
- They shared only 41% of their cited sources.
Practical meaning:
- One screenshot of ChatGPT recommending you, or ignoring you, means very little.
- Measure across a fixed set of questions, repeated over time, and look at the share of answers you appear in.
Finding 5: Reddit, YouTube and review sites are a small slice of citations
For these B2B software questions, Reddit made up about 1% of all citations, YouTube about 1%, and G2/Capterra-style review sites about 1%. Gemini was the exception, with Reddit at 5% of its citations.
That doesn't mean community and review sites don't matter. Models can read a Reddit thread and still cite something else, and reviews influence the articles that do get cited. But for B2B software buying questions, the sources that show up in the answer are overwhelmingly roundups, comparisons and vendor pages.
Finding 6: Problem questions often name no product at all
Some of the highest-volume questions in our set got answers that recommended no software at all. They cited agency blogs and platform documentation instead:
- "How much does SEO cost per month for a small business?" cited agency pricing blogs
- "How do I optimize my website to show up in ChatGPT search?" cited only OpenAI's help docs
- "How should a B2B company do Reddit marketing without getting banned?" cited only Reddit's help pages
These are open slots. Whoever publishes the most useful, specific, data-backed answer has a real chance of becoming the source.
What we'd do with this
- Build a fixed question set (30–70 questions) around real buying decisions: category, comparison, alternatives, cost, use case, industry. Rerun it monthly.
- Track each engine separately. Your ChatGPT, Claude, Gemini and Perplexity results will differ.
- For ChatGPT, make your own product, pricing and comparison pages specific. It cites them directly.
- For Claude, Gemini and Perplexity, earn a place in the third-party roundups they already cite. Open the sources behind the answers you lose and work through that list.
- Answer problem questions with original data. Pricing, benchmarks, before/after results. Generic summaries give a model no reason to pick you.
- Measure share of answers, not single screenshots.
Okara's GEO agent runs this kind of tracking and turns the gaps into fixes: content, technical changes shipped as pull requests, and third-party opportunities. See how it works →
Methodology notes
- 230 answers collected November 7, 2026, US English, web search enabled.
- ChatGPT data is from the live chatgpt.com interface; Claude, Gemini and Perplexity data is from their APIs with web search.
- "Brand overlap" and "domain overlap" are Jaccard similarity between the sets of brands or cited domains in two answers to the same question.
- Page types were classified by URL pattern; roughly 12% of URLs didn't fit a category and are excluded from the table.
- Full question list available on request.