ChatGPT

All figures measured.

#ProductRecommended in
1 aflo 3 of 3
2 MailFlow 3 of 3
3 MailYVO 3 of 3
4 Sentto 3 of 3
5 Uppush 3 of 3
6 Brevo 2 of 3
7 Mailchimp 2 of 3
8 VerticalResponse 2 of 3
9 Wiz 2 of 3

Scanned 12 October 2026. 3 samples from ChatGPT, round 9. 9 products recommended. Methodology v2026-09-29.1.

Perplexity

All figures measured.

#ProductRecommended in
1 Sender 3 of 3†
2 Shopify Messaging 3 of 3†
3 Mailercloud 2 of 3†
4 Shopify Email 2 of 3†

† Corrected 12 October 2026: incomplete_round. Some of this round's samples had failed when it was first published. They were collected afterwards with the same questions and settings, and the figures marked † were recomputed. How corrections work

Scanned 12 October 2026. 3 samples from Perplexity, round 9. 4 products recommended. Methodology v2026-09-29.1.

Counts, not percentages: with three samples a percentage would imply a precision the sample size cannot carry. “2 of 3” means the assistant recommended that product in two of its three answers to this question.

Sources the assistants cited

  • apps.shopify.com 5
  • techradar.com 3
  • shopify.com 3
  • emailwiz.ai 2
  • sender.net 2
  • reddit.com 2
  • layers.to 1
  • trymailflow.com 1
  • paslaugos.lt 1
  • mailyvo.com 1
  • prtimes.jp 1
  • afloai.com 1

Scanned 12 October 2026. 6 samples across ChatGPT and Perplexity, round 9. Methodology v2026-09-29.1.