Envoyra AI Recommendation Report, Fall 2026

At the top, the shelf is steady. HubSpot was recommended in 305 of 368 answers about CRM for small business, and its weekly counts barely moved: 77, 76, 75 and 77. Intercom was named in 62 of 96 customer support answers in every one of the four weeks.

Below the top, the answers keep moving. We tracked 680 pairs of one question and one assistant. Only 144 of 680 pairs gave the same list of brands in all four weeks. The rest changed at least once: a brand appeared one week and was gone the next, or a new name joined.

Checked against outside research. This held up. SparkToro, working with the AI-tracking company Gumshoe, asked ChatGPT, Claude and Google's AI the same buyer questions many times. The same list of brands came back fewer than 1 time in 100, yet a handful of leading brands kept turning up.^[1][2][3] Their study covered three AI tools; ours covers four.

Separately, we asked each assistant about 30 of these brands by name, 24 times per brand. An answer counted as recognized when it correctly said what the brand does, in its category. Then we set that next to how often each brand was recommended when buyers asked about the category without naming anyone.

The two don't line up.

Known, but rarely recommended. Every assistant knew Keap: 24 of 24 answers described it correctly. Yet Keap was recommended in only 3 of 400 CRM answers.² Close was recognized in 24 of 24 answers and recommended in 16 of 352.

Recommended, but not recognized by name. Some brands appear on the shelf when buyers ask about the category, yet the same assistants don't place them when asked by name.

Many of the misses were about a different company with the same name. Asked about Fathom, several answers described a financial-reporting company also called Fathom.^[6] Asked about Profound, some described Profound UI, a developer tool from a company called Profound Logic.^[7] Asked about Granola, some described an older energy-saving utility for computers, also called Granola.^[8] These are real companies and products, past or present, so the assistants weren't inventing anything. They picked the wrong one. Otterly.AI appears to be confused with Otter.ai.^[9] Some answers about Peec AI described products that don't match it.

Checked against outside research. The same-name pattern held up. Academic work shows language models often mix up different things that share a name, and answer inconsistently when a name could mean more than one thing (AmbigDocs, 2024; Sedova et al., 2024).^[11][12] We found no published study that sets recognition by name against recommendation, so this comparison appears to be new.

² Among the 30 brands tested: Keap was acquired by Thryv (announced October 2024).^[13] Forethought is part of Zendesk (since March 2026).^[14] Brevo was formerly Sendinblue.^[15]

• Questions. Buyer-style questions about each category, asked the same way every week through each assistant's API. We don't publish the questions.
• Assistants. ChatGPT (gpt-4o with web search), Claude (Sonnet 4.6, no web search; reliable knowledge cutoff August 2025^[16]), Gemini (2.5 Flash, no web search; knowledge cutoff January 2025^[10]) and Perplexity (sonar-pro, with web search).
• Weekly runs. Four Sunday runs from Sep 13 to Oct 4, 2026, 170 questions per assistant per run, 2,720 answers in total.
• Counting. Each answer counts once per brand. A question that names a brand is left out of that brand's count, so the totals differ slightly from brand to brand.
• Name-recognition test. Short name-only questions about each brand on each assistant, repeated over three days from Oct 7 to Oct 9, 2026, giving 24 answers per brand. This test was set up differently from the weekly questions, so its results are reported separately. Answers were labelled against a written rubric. A second, independent check reviewed the doubtful labels, and Envoyra founder Golara Serio decided the disagreements. Recognition and recommendation are reported side by side, never combined.
• Categories. Each brand's category was checked against the company's own website on Oct 9, 2026 (listed under Sources).^[17]

Limits

• These are observed outputs, not hidden reasoning. We report what the assistants said, not why.
• Results hold for these assistants, these questions and this window.
• The recognition test is small: 6 answers per brand per assistant.
• In this data, only Perplexity's sources were recorded, so claims about sources in this edition cover Perplexity only. Later editions will add sources from ChatGPT and the other assistants that share them.
• The exact model version each provider served was recorded from Oct 4 onward. For the earlier runs we know which model we asked for, not the exact version that answered.
• We don't rank vendors.