Latent Space published an AEO Tracker that measures how many sources each frontier model cites

Latent Space published the Frontier AEO Tracker on 7 September, a hosted tool at aeo.latent.space that asks seven frontier models the same product questions and records the sources each one cites. Latent Space calls this Answer Engine Optimisation, the counterpart to search engine optimisation for the AI answers that increasingly replace a results page.

What the tracker actually runs

Latent Space says the tracker covers 161 product categories, from coding tools to databases to angel investors, and asks each category through 6 prompt variations. Seven frontier models answer: Claude Code with Fable and Opus, Codex with Sol and Astra, Cursor with Grok, Muse Code, and Devin's SWE-1.7. Latent Space says Gemini's Antigravity, GLM's Zcode and DeepSeek's DeepCode could not be included because of technical limits. Answer extraction runs on Astra, and every prompt and answer pair is inspectable on the site.

The number the site leads with

Latent Space reports a wide gap in how many sources each model brings back for the same question. The median counts it publishes are 5 for Astra, 9 for Sol, 11 for Opus, and 15 for Fable. If those numbers hold up under repeat runs, they say that being cited by Fable is a different task from being cited by Astra: one is a shortlist, the other is a reading list, and a page that reaches the reading list may never reach the shortlist.

Latent Space adds one qualitative finding worth flagging: Astra's picks are more consistent across paraphrased versions of the same question than Sol's or Fable's. The company argues this makes AEO measurable at all, because the randomness that used to swamp any measurement is falling.

The claim under the claim

The tracker itself is Latent Space's own work, so the numbers it publishes are its measurements, not a third party's. Latent Space is direct about what the tool is: an analysis project, not an ongoing service. Update cadence is not stated, and the underlying prompts are frozen at the version the site publishes. Rerunning the same categories a month from now, against the same models at a later revision, would produce different numbers.

The category list is also a choice. "Angel investors" and "databases" answer to different buying processes, and a single ranking that mixes them will hide inside-category noise. The value of the tool is in the per-category breakdowns and the inspectable prompts, more than in any headline league table.

A team that sells a developer tool now has somewhere to point when it asks why it is not cited. Fable read fifteen pages to answer this question; if this team was not on any of the fifteen, that is the gap to close. Astra read five, so the shortlist test is stricter, and getting onto it is the harder work.

Source

The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it), Latent Space, 7 September 2026.