AINEWS
A new public tracker asks seven AI models the same buyer questions across 161 software categories, including which CRM, IDE, and note taking app to use. The seven disagree more than you'd expect.
When you ask an AI assistant which CRM, IDE, or note-taking app to use, the answer is no longer a search result. It's a single recommendation, usually confident, and it varies depending on which AI you asked.
That shift has given rise to a new discipline: AEO, or answer-engine optimization. Where SEO tried to game Google, AEO tries to game the AI assistants that now act as the first stop for software-buying questions. A new public tracker from Latent Space, built on AmplifyingAI's earlier "What Claude Code Actually Chooses" study, puts a number on that layer for the first time.
The methodology is straightforward. Six prompt variations, from blunt best-of to casual phrasing, were posed to seven frontier AI systems across 161 software categories, with web search enabled. The full prompt-and-answer pairs are hosted at aeo.latent.space. Each answer was extracted and scored by Astra on a proprietary scale: +60 for a first-choice recommendation, +25 for an alternative, +15 for a substantive mention, -25 for a scenario objection, -60 for a broad objection. The total is divided by 60 and expressed as a percentage, so a model that consistently leads a category with no real pushback lands near 100%.
The seven models tested are Fable (Claude), Opus (GPT), Grok, Sol (Cursor), Astra, Muse (Codex), and SWE-1.7 (Devin). Three more were attempted: Gemini/Antigravity, GLM/Zcode, and DeepSeek/DeepCode. They were dropped because of rate limits and errors. The exclusion is a real coverage gap, not a footnote.
Where the seven agree
Out of 61 categories with complete coverage across all six prompt variations, 28 have a universal primary choice. The same product wins no matter how the question is phrased: pnpm for JavaScript package management, Sentry for error monitoring, Zoom Workplace for video conferencing, PostgreSQL for relational databases, Vitest for unit testing, Convert for A/B testing. These are the table-stakes picks the AI layer has already standardized on.
Where they disagree
In 33 of 121 comparable categories, Sol and Astra land on different leaders, with each model repeating its own pick across all six prompt variations. Forty-one category leaders finish within 10 points of the runner-up. In the AI app builder category, for example, Lovable leads with 52.4% and Replit trails at 50%. That four-point gap is the AI layer in real time: a buyer asking "what's the best AI app builder?" gets a coin-flip answer depending on which model they trust.
The median count of first-choice picks across the 161 categories tells its own story: Fable (Claude) leads with 15, Opus (GPT) with 11, Sol (Cursor) with 9, Astra with 5. Astra is far less likely to change its mind when you lightly paraphrase the question, which means its recommendations are more consistent but also narrower.
The soft home-bias pattern
Each model has a tendency to recommend its own lab's product. Fable and Opus pick Claude Code. Sol and Astra pick Codex. Grok picks Cursor. Muse picks Muse Code. SWE-1.7 picks Devin. The Latent Space authors describe this as a soft bias, and it shows up in the per-category leaderboard. Whether it's a training-data artifact, an RLHF preference, or a deliberate vendor nudge is unclear; the tracker doesn't try to settle that. The signal is that the AI recommendation layer is not neutral, and the bias is uneven across categories.
Visibility without selection
Some products are everywhere and nowhere. npm is mentioned in 42 of 42 measurements but never gets a first-choice nod. Jest does the same: 42 of 42, zero first choices. Notion follows at 41 of 42. These are the entrenched defaults, known by every model, picked by none as "the best." For a buyer, that's useful: visibility is not endorsement.
What to do with the map
The tracker's real value isn't the leaderboard. It's the disagreement layer. When Sol and Astra split on a category, a buyer can treat that as a triangulation signal: ask both, weigh the answer against their own stack, and stop trusting a single model's first pick. The same pattern works for vendors. If your product is consistently visible but never selected, the AI layer treats you as the safe default. If you're invisible, the AI layer is recommending against you, often without saying so.
The full prompt-and-answer pairs are public at aeo.latent.space, so anyone can read their own category and see how the seven models actually answer. That's the constructive hinge: the methodology is transparent, the data is downloadable, and the bias is at least legible. AEO is shaping up as a market layer, and the first map of it is on the table.