We Measured Our Own AI-Search Invisibility
Last verified: 2026-09-25. Ten valid snapshots (one more discarded as invalid). Same 10 queries. Same model (sonar). This is not six months of data and it is not a ranking guarantee.
E-GEO is an open-source GEO/AEO toolkit. For a month we optimized the repo (topics, description, README) and still did not appear when Perplexity was asked the category question: What are the best open-source generative engine optimization (GEO) tools on GitHub in 2026?
That query was false on every snapshot for six weeks. On 2026-09-21 it turned true for the first time — one snapshot, not yet a trend. This page is the measurement log, not a success story.
What we measured
Section titled “What we measured”A fixed set of 10 queries, weekly, via Perplexity Sonar. A mention is a binary: the answer text referred to E-GEO / eGEOagents. Branded queries are not treated as category wins.
| Date | Mentions | Head query (best OSS GEO tools on GitHub 2026) | Generic (non-branded) hits |
|---|---|---|---|
| 2026-08-05 | 2/10 | false | none |
| 2026-08-09 | 2/10 | false | none |
| 2026-08-10 | 3/10 | false | GEO evaluation harness open source |
| 2026-08-11 | 3/10 | false | same |
| 2026-08-17 | 4/10 | false | harness + open source tool to rank in ChatGPT and Perplexity |
| 2026-08-24 | 3/10 | false | harness only (the extra 2026-08-17 hit did not hold) |
| 2026-08-31 | 3/10 | false | harness only |
| 2026-09-07 | — | — | discarded: every answer came back empty (API failure); the script now marks such runs invalid instead of recording zeros |
| 2026-09-11 | 3/10 | false | harness only |
| 2026-09-14 | 3/10 | false | harness only |
| 2026-09-21 | 5/10 | true | head query + open source AEO tools + harness |
The two branded queries (eGEOagents, eGEOagents GitHub) were true from day one. They do not count as category visibility.
Single-run answers vary. The 5/10 on 2026-09-21 is one sample. Two of those five are branded, so the category picture is 3 generic hits in one run. In that run, the sources Perplexity cited for the head query included LibHunt, GitHub and egeoagents.com itself.
What actually changed off-repo
Section titled “What actually changed off-repo”On-repo metadata was already exhausted by 2026-08-05. After that we shipped external surfaces:
| Surface | When | Status |
|---|---|---|
| Docs site egeoagents.com | 2026-08 | live |
| LibHunt listing | 2026-08-09 | live, auto-approved |
| GitHub Sponsors + a $199 GEO audit | 2026-08 | live |
egeo on PyPI |
2026-08-31 | live (pip install egeo) |
GitHub stars moved 147 → 168 in August (194 by 2026-09-25). The head query did not move for six weeks. Stars were never the bottleneck — projects with 17–37 stars already appeared in that answer because they were in the sources Perplexity cites.
What we still cannot claim
Section titled “What we still cannot claim”- One true head-query snapshot is not a position. We need it to hold before claiming we are in that answer; geo-optimizer-skill, geo-lint, xanlens, open-geo, GEO-optim/GEO and the izak-fisher / amplifying-ai lists appeared in it more consistently.
- The 2026-08-17 extra generic mention regressed the following week. The 2026-09-21 hits may too.
- Several things shipped between 2026-08-31 and 2026-09-21 (this case-study page on 2026-09-01, PyPI releases 2.0.0 and 2.0.1). We cannot attribute the 2026-09-21 change to any single one.
- arXiv:2511.20867 is a preprint, not a peer-reviewed paper. The peer-reviewed GEO anchor is Aggarwal et al., KDD 2024 (arXiv:2311.09735).
- E-GEO consumes MCP servers as a client. It is not an MCP server.
- Indexation, crawl, and citation are different events. We do not claim Google or Perplexity will index or cite a URL because we asked.
Method, so you can repeat it
Section titled “Method, so you can repeat it”- Freeze 8–10 queries. Do not edit them week to week.
- Ask the same engine. Record mention yes/no, competitors named, and cited domains.
- Treat cited domains as the distribution list. In this category that was LibHunt first, then awesome-lists and dated comparison pages.
- Do not read a branded hit as a category win.
- Re-run a one-week spike before acting on it.
The playbook distilled from this log is in How to get cited by Perplexity. The research lineage is on /research/.
Next measurement
Section titled “Next measurement”The next weekly snapshot is 2026-09-28. Gate A in our public plan is ≥3/10 sustained or ≥300 GitHub stars. Numerically, ≥3/10 has held on the last five valid snapshots — but two of those three-plus hits are branded every time, so we do not count the gate as met until the generic hits hold. Stars: 194 of 300.
In short: a GEO tool with a polished README still does not exist to Perplexity until curator sources name it. We measured that on ourselves: six weeks invisible on the head query, then one true snapshot once the curator sources and our own site were in place. Install egeo or read the source.