What Perplexity cites when Amazon sellers ask: 6,946 answers, April–November 2026
Between May and November 2026 Perplexity moved its sources for Amazon seller questions from independent sites to Amazon's own pages. On the same 64 questions, answers citing sell.amazon.com rose from 33.2% in September to 52.7% in November, while Reddit fell from 34.2% to 3.8%, novadata.io from 53.0% to 21.7% and helium10.com from 21.4% to 5.6%. Across all 5,230 answers to questions that name no tool, Helium 10 was the brand named most often (52.1% of answers), followed by Jungle Scout (39.7%) and Sellerboard (24.8%).
Key findings
- Amazon's own pages took over. The share of answers to unbranded questions citing at least one Amazon-owned page (sell.amazon., sellercentral., globalselling.amazon.com and similar) went from 58.9% in May to 84.0% in November.
- Reddit and YouTube almost disappeared as sources. Reddit fell from 48.8% of answers in August to 3.5% in November; YouTube from 86.4% in May to 2.1%.
- Independent tool blogs lost the most. On identical questions, September → November: novadata.io 53.0% → 21.7%, sellersprite.com 30.7% → 9.9%, sellerassistant.app 25.3% → 9.9%, helium10.com 21.4% → 5.6%, agentxray.ai 36.3% → 3.8%.
- Being cited is not being named. Over the whole period agentxray.ai was cited as a source in 21.0% of unbranded answers, but AgentXray was named as a tool in only 4.7%. A page can feed an answer without the answer mentioning who wrote it.
- Brand visibility fell with the blogs. On identical questions, Helium 10 was named in 58.8% of answers in September and 40.4% in November; Jungle Scout 43.6% → 34.5%; Sellerboard held at 32.1% → 30.3%.
The October shift, on the same 64 questions
The question bank changed over time (see Methodology), so this table compares only the 64 unbranded questions asked in both September (1,516 answers) and November 1–9 (446 answers). Each figure is the share of answers that cite the domain at least once.
| Source | September 2026 | November 2026 |
|---|---|---|
| sell.amazon.com | 33.2% | 52.7% |
| sellercentral.amazon.de | 15.5% | 36.8% |
| sellercentral.amazon.com | 19.9% | 30.5% |
| revenuegeeks.com | 38.7% | 23.5% |
| novadata.io | 53.0% | 21.7% |
| sell.amazon.de | 15.0% | 21.5% |
| sellercentral.amazon.es | 0.4% | 16.6% |
| globalselling.amazon.com | 0.4% | 15.7% |
| github.com | 2.1% | 14.3% |
| sellersprite.com | 30.7% | 9.9% |
| sellerassistant.app | 25.3% | 9.9% |
| helium10.com | 21.4% | 5.6% |
| agentxray.ai | 36.3% | 3.8% |
| reddit.com | 34.2% | 3.8% |
The number of sources per answer did not drop — across all unbranded answers it averaged 17.6 in September and 15.4 in November — so the change is in which sources are chosen, not in how many.
Month by month (unbranded questions)
| Month | Answers | Questions | Cites an Amazon page | Cites Reddit | Cites YouTube | Cites agentxray.ai | Names AgentXray |
|---|---|---|---|---|---|---|---|
| May | 817 | 64 | 58.9% | 45.0% | 86.4% | 7.1% | 3.2% |
| August | 776 | 64 | 60.8% | 48.8% | 15.3% | 32.2% | 9.8% |
| September | 1,569 | 64 | 69.6% | 33.5% | 4.8% | 35.6% | 8.2% |
| October | 1,568 | 64 | 86.0% | 11.5% | 2.8% | 13.5% | 1.0% |
| November 1–9 | 482 | 82 | 84.0% | 3.5% | 2.1% | 3.5% | 0.2% |
No answers were collected in June and July. April (18 answers, the first test run) is left out.
Which tools get named
Share of the 5,230 unbranded answers that name each brand, April–November. Only the nine brands the tracker looks for are counted.
| Brand | Answers naming it | Share |
|---|---|---|
| Helium 10 | 2,726 | 52.1% |
| Jungle Scout | 2,075 | 39.7% |
| Sellerboard | 1,299 | 24.8% |
| Keepa | 1,224 | 23.4% |
| AMZScout | 1,031 | 19.7% |
| SellerApp | 429 | 8.2% |
| ZonGuru | 296 | 5.7% |
| Viral Launch | 82 | 1.6% |
| DataDive | 22 | 0.4% |
The most cited sources over the whole period
Share of the 5,230 unbranded answers citing each domain at least once. 1,515 distinct domains were cited in total.
| Domain | Answers | Share |
|---|---|---|
| novadata.io | 2,038 | 39.0% |
| sell.amazon.com | 1,946 | 37.2% |
| reddit.com | 1,469 | 28.1% |
| revenuegeeks.com | 1,344 | 25.7% |
| sellercentral.amazon.com | 1,243 | 23.8% |
| sellersprite.com | 1,175 | 22.5% |
| sellerassistant.app | 1,098 | 21.0% |
| agentxray.ai | 1,096 | 21.0% |
| helium10.com | 1,023 | 19.6% |
| youtube.com | 969 | 18.5% |
| sellercentral-europe.amazon.com | 903 | 17.3% |
| sellercentral.amazon.de | 864 | 16.5% |
| sellerlogic.com | 817 | 15.6% |
| sell.amazon.de | 776 | 14.8% |
| osome.com | 692 | 13.2% |
The full list of the top 30 domains, the brand counts and the monthly shares are in the CSV download.
The same pattern on Microsoft Copilot
Copilot is measured differently: Bing Webmaster Tools reports how often a site's pages were used as a source in Copilot answers. For agentxray.ai it reported 557 citations over August–November 2026, 308 of them on one page. Daily citations ran at 8–15 until late October and 0–3 afterwards. Part of that fall is our own doing — we unpublished that page on 30 September and Copilot stopped citing it about three weeks later — so Copilot's numbers cannot separate a platform shift from our change. We report them for completeness, not as evidence of the Perplexity shift.
What this means
- For sellers: answers about fees, FBA rules and marketplace requirements now lean on Amazon's own documentation. That is good for accuracy on policy questions and worse for comparisons between tools, where Amazon's pages say nothing.
- For tool makers and publishers: a page that was a top source in September could be a rare one by November without any change on the site. Visibility in AI answers has to be measured over time, on the same questions, or a platform change will look like a site problem.
- For anyone measuring GEO: count citations and named mentions separately. Over this period the two differed by more than four times for us (21.0% against 4.7%).
Methodology
- Engine: Perplexity's
sonarmodel through its API (via OpenRouter), web search on. The API model is not guaranteed to match the answers shown in the perplexity.ai app. - Window: 27 April – 9 November 2026, with no collection in June and July. 6,946 successful answers to 104 distinct questions.
- Questions: written by AgentXray for its own visibility tracking and tilted towards European sellers, profit and operations. 5,230 answers came from questions that name no tool ("category" and "problem" questions); 1,716 from questions that name a brand, reported separately and excluded from the tables above. The bank changed over time: 64 unbranded questions from May to October, 82 from 7 November. The September–November comparison uses only questions present in both periods.
- Counting: a domain is counted once per answer, however many of its URLs the answer lists. "Amazon pages" means any domain under amazon.* (including sellercentral, sell and globalselling subdomains). Brand counts use a fixed list of nine brands; other tools named in answers were not counted. "Names AgentXray" uses the tracker's validated mention: the name appears in an Amazon-tooling context, not merely as a string.
- Reproduce: every number on this page comes from
scripts/geo/ai_citation_study.pyrun on 9 November 2026.
Limitations
- One engine, one company's question set. The results describe these 104 questions, not Amazon seller questions in general.
- AgentXray sells Amazon seller software and appears in the data. We have reported our own numbers as they came out, including the fall.
- November covers nine days.
Published by AgentXray Research (Avanta Global EOOD). Corrections: [email protected].