How AI assistants decide which brands to recommend (2026 data)

July 1, 2026 · 8 min · Spike research

Ask ChatGPT "what's the best CRM for a 10-person startup" and it answers in four seconds with three names. Somewhere, a founder just lost a deal they never knew existed. This post is about how those three names get picked — based on what Spike's visibility intelligence observes across thousands of AI answers every week.

The short version

AI assistants recommend brands through two distinct pathways, and most companies optimize for neither:

1. Parametric memory — what the model learned in training. Slow to change, heavily weighted toward brands with years of consistent public description.
2. Retrieval — what the model reads right now when it searches the web before answering. Changes weekly. This is where challengers win.

Perplexity and ChatGPT-with-search are retrieval-first. Gemini blends both. Open models like Llama answer almost purely from memory. A visibility strategy that ignores either pathway covers half the market.

What the data shows

Across the thousands of AI answers we analyzed in B2B SaaS categories in June 2026, a few patterns repeat with almost boring consistency:

• Category leaders dominate memory answers. In prompts like "best X", the top brand by G2 review count appeared in 91% of memory-only answers. The #4 brand by reviews appeared in under 20%.
• Retrieval answers cite a narrow source set. Review aggregators (G2, Capterra), Reddit threads, and "best X in 2026" listicles account for the large majority of citations we see. Brands' own sites are cited far less — usually only when structured data or a comparison page exists to quote.
• Rank decays fast. Being mentioned third instead of first costs roughly a third of the click-equity; below rank five, mentions barely matter.
• Freshness is a tiebreaker. When two brands are otherwise comparable, engines consistently prefer the one whose pages carry recent dateModified signals and visibly updated content.

The mechanics, engine by engine

Memory-first engines (Llama, Mistral, base Gemini) compress the internet's consensus about you. They can't be persuaded quickly — they reflect what review sites, documentation, comparison pages, and news said about you over years. The lever here is consistency: the same one-line description of what you are, everywhere it can be read.

Retrieval engines (Perplexity, ChatGPT with browsing, Copilot) do a live search, read the top handful of results, and synthesize. Their answers are only as favorable as the pages they retrieve. If the top results for best <your category> don't include you, the answer won't either. The lever here is owning retrievable surfaces: comparison pages, FAQ pages that literally answer buyer questions, and presence on the review sites engines trust.

Why brands are invisible

When we scan a SaaS that scores under 20/100, it is almost always the same five gaps — not "bad marketing":

1. No structured data, so engines have no machine-readable facts to quote.
2. No llms.txt, so crawlers reconstruct the product from nav links.
3. robots.txt blocking GPTBot or CCBot (often left over from a 2023 decision).
4. No page that answers a buyer question in quotable form.
5. No comparison content, ceding every "X vs Y" prompt to competitors.

Every one of these is fixable in days, not quarters. We wrote up the complete checklist in the 7 fixes that make AI cite you.

What to do with this

Treat AI answers as a channel with a feedback loop: measure where you appear, ship the fixes, re-measure weekly. Retrieval answers move within weeks; memory answers follow over quarters as the corpus about you improves. The compounding starts the day engines can read you properly.

Spike runs this loop automatically — scan your site free to see your baseline across six engines.