I Built an Open-Source AI Visibility Tool, Tested It on Myself, and Scored 0/16 (OpenGEO)

Somebody in your category is typing this into ChatGPT right now: what are the best open-source alternatives to Otterly AI and Profound for AI search visibility? I ran that exact prompt across ChatGPT and Google AI Overviews and got back tidy shortlists of five to six projects — Elmo, Limelit Open, GEO/AEO Tracker, OneGlanse, Gego, Searchstack AEO, GEO Optimizer, Canonry — each with a crisp summary, a repository or product link, and a sharp best for framing. Two sections later the model had sorted them into a comparison table and named a first pick. A buyer who reads that answer is done shopping. They never scroll to page two.

Our own open-source project — OpenGEO — came back at 0 out of 16 captured answers across the eight core English prompts we tested on ChatGPT and Google AI Overviews. Not ranked fifth — completely absent on category alternatives, absent on workflow guides, and even when we asked directly What is OpenGEO and how does it help track AI search visibility locally?, both engines either attributed the name to an unrelated Pupok462/open-geo repository or copied "Google Lighthouse for AI search engines" straight off another project's card. The uncomfortable realization isn't that everyone else built a better product. It's that AI search builds shortlists exclusively from the pages already legible to it, and our repository wasn't even in that candidate pool.

Our baseline in OpenGEO before rewriting docs — 0 out of 16 captured ChatGPT and Google AI answers mentioned cooleryu/opengeo.

Here is the practical audit-and-rewrite loop that gets you into that pool, and you can run the first round in an evening: ask the exact questions buyers actually type, store every verbatim answer and cited source, inspect the top three winner pages side by side, and rewrite your own front page and onboarding docs to close the structural gaps. Once your changes ship, re-run the identical prompts and keep both rounds on one timeline.

What does an AI answer look like when it recommends your competitor?

It reads like a procurement brief with sourcing baked in. Look closely at what showed up when we opened Question #2 (open-source alternatives to Otterly AI and Profound) inside our captured evidence view:

Capturing the exact question alongside the full answer and every competing tool named instead of you.

Every candidate that made the cut carried four consistent building blocks: a plain category definition in line one, a clear license and deployment signal (Self-hosted · GPL-3.0 or MIT · SQLite), a bulleted capability list using buyer language, and an implicit or explicit best for sentence. Beneath the prose sat GitHub repositories and concise product homepages. You aren't competing with a competitor's roadmap in that moment — you're competing with how clearly their landing page answers the prompt.

How do I capture real answers without buying an enterprise tracker?

Start with a fixed list of real buyer phrasing, ask across the engines your users trust, and keep the full raw text — never summaries. Four fields per record decide whether later comparisons actually mean anything:

• The exact prompt word for word. Best GEO tools and how can indie SaaS teams run local GEO checks without enterprise subscriptions trigger completely different retrieval sets. Lock the prompt once and reuse it unchanged.
• Engine, entry point, and date. ChatGPT logged-in web answers, Google AI Overviews, and developer API endpoints read different source snapshots on the same day. Never diff an API response against a consumer browser capture.
• Verbatim answer text. If your product name or disambiguating context isn't inside the captured text, you were skipped — full stop.
• Every named competitor and external link. These URLs are your real content competition for the next sprint.

Doing this in spreadsheets works for three prompts and falls apart once you track eight prompts across two platforms before and after edits. That bookkeeping annoyance is why we keep everything local inside OpenGEO: it runs your question set inside your own logged-in browser session at a calm human pace (pausing and handing control right back to you if Google pops a verification challenge), or takes full manual copy/pastes from an incognito tab, and keeps mentions and official-site citations audited separately on one screen.

flowchart LR
    A["Ask the exact question a buyer types"] --> B["Save the full answer and every cited URL"]
    B --> C["Open the winner pages and map their structure"]
    C --> D["Rewrite your front page and guides to answer that gap"]
    D --> E["Publish, then re-ask the identical question"]
    E --> F["Compare before and after on one timeline"]

Which pages actually got cited, and what do they share?

Before changing a single paragraph on your own site, probe the external pages that already won the prompt. Across the winners we inspected (getgego.org, elmohq/elmo, limelit-co/open), five patterns showed up over and over again:

1. A category definition in sentence one. Not marketing copy. Plain grammar: subject, category, deployment model, primary job (Self-hosted GPL-3.0 CLI + dashboard or Open-source, local-first AI search visibility workbench).
2. H2 headings phrased around buyer concerns. Look at the live headings we pulled from getgego.org: Know how AI sees your brand, One tracker for every major LLM, From SEO research to competitive GEO. Models extract section headings verbatim when assembling "how it works" summaries.
3. Explicit disambiguation if a similar name exists. If another repository or extension shares your name, state who you are and who you are not directly on your canonical README so retrieval models stop merging the two entities.
4. Honest best-for and not-for boundaries. Saying hosted SaaS tools are better if you want zero setup and weekly hands-off reporting builds credibility and gives comparison prompts clean material to quote.
5. Real product screenshots with descriptive alt text. Explaining an abstract workflow with three tables loses both developers and extractors; showing what the evidence view and draft diff look like anchors every claim.

What you borrow from winners is their outline clarity and directness — never their wording, claims, or benchmarks.

How do open-source and local GEO tools differ from paid SaaS monitors?

They answer different halves of the problem. Hosted monitors tell a marketing team whether share-of-voice moved last Tuesday. Self-hosted and local tools keep the raw evidence in your hands so engineers and docs owners know which exact file to ship next week.

Two practical realities belong right next to that table: self-hosted still incurs API or model billing when you query online or generate drafts, and if all you want is hands-off weekly keyword reporting across ten engines, a paid SaaS monitor genuinely saves time. Choose the tool built for the bottleneck you actually have.

Turning cited-page gaps into drafts you actually ship

Finding a gap isn't the finish line; staring at an empty Markdown file next Tuesday is where most teams stall. Inside OpenGEO's optimization studio, once we pick the missing questions (0/16 baseline across category and brand prompts), the workbench probes the cited winner pages first and lays out a small set of complementary deliverables rather than three copies of the same pitch: a clearer repository README, a concrete Getting Started guide, and a public field teardown like this post.

Optimization Studio — review each draft against its target prompts, inspect what changed line by line, and adopt or export one piece at a time.

Every project profile, captured answer, and API key stays on your own disk, so browsing saved evidence costs nothing and nothing ever pushes online without you choosing to adopt or copy it out.

What should I rewrite first, and how do I verify it worked?

Focus on your highest-leverage canonical page before scattering edits everywhere else:

1. Rewrite sentence one of your README or homepage to state category + deployment + core loop in plain nouns and verbs.
2. Turn major H2 headings into concrete buyer questions, especially comparison and setup questions (How it differs, Supported platforms, How it works).
3. Add one explicit disambiguation line near the footer (or quick intro if actively confused) naming the canonical repo URL and what unrelated projects it isn't affiliated with.
4. Ship one workflow guide or comparison teardown backed by real captures and real UI screens instead of generic marketing claims.

After shipping, give search crawlers and model indexes several days to two weeks — re-testing ten minutes later just hits stale caches. Then re-run the exact same prompt on the exact same browser or API entry point and compare before vs. after side by side: first look for your domain appearing inside cited links (that unlocks the candidate pool), then verify whether your product's description stopped getting mixed up with lookalikes.

Try a ten-minute audit tonight

Pick one question a real customer would ask before buying your product, run it in a clean ChatGPT or Google session, and save the full text and every link underneath. Count how many of those cited URLs point back to you. That single honest baseline is worth more than any guesswork dashboard.

If you want to manage that whole loop — local browser capture, winner breakdowns, draft diffs, and side-by-side retesting — on your own machine, OpenGEO is open source at https://github.com/cooleryu/opengeo. Drop the repo URL into the AI coding assistant you already use (Claude Code, Cursor, Codex, or Antigravity) and ask it to get the local workbench running so you can inspect your own first eight questions tonight.