State of AI Search 2026: We Asked 5 AI Engines to Recommend Software

Published July 2026 · 11 min read

We took 10 ordinary buying questions — “what’s the best CRM?”, “what’s the best password manager?” — and ran each one three times through five grounded AI engines: ChatGPT, Claude, Gemini, Perplexity, and Grok. That’s 150 real answers. Then we counted every brand each engine named. The result is a clear, and slightly unnerving, picture of who AI recommends — and who it has quietly erased.

Three Findings, Up Front

If you read nothing else, read these:

• Every category has a locked shortlist. In all 10 categories, at least one brand was named in all 15 runs (five engines, three times each). The engines converge hard on the same handful of leaders. On average, the top three brands captured 66% of all mentions in a category.
• The leaders are stable; the challengers flicker. Overall, 25% of brand appearances were inconsistent across three identical runs — but that noise is concentrated in the middle tier. The top brands showed up every single time. Ask once and you might miss a challenger; ask three times and the leaders never move.
• Well-known brands can be completely invisible. Eleven established products were never named once across 150 answers. Notion did not appear for engineering project management. NetSuite did not appear for accounting. Carrd did not appear for website builders. Reputation in the market is not the same as visibility in AI.

Who AI Recommends, by Category

Here is the full board. “Always named” means the brand appeared in all 15 runs. “Sometimes named” shows the share of the 15 runs a brand appeared in. “Never named” means zero mentions across every engine and every run.

The Consensus Is Real — and It’s a Moat

The striking thing isn’t that AI has opinions. It’s how much the five engines agree. Ask any of them for the best CRM and you get Salesforce, HubSpot, and Pipedrive — every time, from every model. Best help desk? Zendesk and Freshdesk, universally. Best video conferencing? Zoom, Google Meet, Microsoft Teams, no exceptions.

This convergence is a moat for whoever is inside it. When five different AI systems, trained by five different companies, all reach for the same two or three names, that shortlist becomes self-reinforcing: they cite the sources that mention those brands, which makes those brands the default answer, which generates more content mentioning them. If you’re on the list, you compound. If you’re not, you have to break in from outside.

Being Famous Isn’t Being Visible

The most useful column in that table is the last one. These are not obscure products — they are brands with real customers and real market share that AI simply did not surface:

• Notion — zero mentions for “best project management tool for a software engineering team.” The engines read “engineering” and reached for Jira and Linear; Notion’s enormous general-purpose brand didn’t register for the specific query.
• NetSuite, Kustomer, Signeasy, Carrd, Cloudflare Web Analytics, Freshsales, Insightly, Nutshell, Basecamp, Smartsheet — all named zero times in their categories.

And plenty of household names were only partly visible. Mailchimp, arguably the most recognized email brand on earth, was named in only 80% of runs — it missed one in five. LastPass appeared in just a third. GoDaddy, despite blanketing the world in advertising, showed up in 7% of website-builder answers. The lesson is blunt: the marketing budget that wins mindshare with humans does not automatically win a citation from a model.

Why You Can’t Trust a Single Check

A quarter of all brand appearances were inconsistent between identical runs. Same question, same model, same day — different answer. If you had checked ChatGPT once and seen your brand, you might have relaxed; if you’d checked once and missed it, you might have panicked. Both would be wrong, because one query is a coin flip on the challengers.

The signal only appears with repetition. Run each question several times across several engines and a stable picture emerges: the locked leaders, the flickering middle, the invisible tail. That’s the difference between an anecdote and a measurement, and it’s the whole reason a real visibility check runs prompts in triplicate rather than once.

What This Means for Your AEO Strategy

Four takeaways if you want to move from the invisible tail toward the locked shortlist:

• Know which bucket you’re in. Always-named, sometimes-named, or never-named is the only diagnosis that matters, and you can only know it by measuring across engines and repeated runs.
• Specificity beats fame. Notion lost the engineering query it could plausibly win because the model associates the category with Jira and Linear. Win the specific phrasing your buyers use, not just general awareness.
• Get into the sources the engines read. Every category leader shares a trait: it’s everywhere in the review sites and listicles models cite. We break down exactly which domains in which sources AI engines cite.
• Re-measure after you ship. Because answers move, you’ll only know your content worked by re-running the same prompts and watching the rate climb. See how to get mentioned in ChatGPT for the tactics.

Methodology

We want this to be reproducible, so here’s exactly what we did and where it’s limited.

• Scope: 10 software categories, one buying question each, run in July 2026.
• Engines: five grounded (web-search-enabled) models — OpenAI gpt-5.4-mini, Anthropic claude-haiku-4-5, Google gemini-3.5-flash, Perplexity sonar, xAI grok-4.3.
• Runs: three samples per question per engine = 150 answers. 149 of 150 returned grounded web sources.
• Counting: a brand counts as “named” when its name or a known alias appears in the answer text. We deliberately excluded candidate brands whose names are common English words (to avoid false matches) and used full product names where needed.
• Limits: we queried grounded provider APIs, not the consumer apps, which can use different model versions and personalization. Ten categories and three runs is a directional sample, not a census. Substring matching can miss an oblique reference or, rarely, over-count. Treat the numbers as a reproducible benchmark, not gospel.
• Tooling and cost: the whole study ran on the open-source openllmrank CLI and cost $6.01 in API calls.

Frequently Asked Questions