GeoLook: An Open Source GEO Tool That Checks If AI Actually Cites You

Most teams checking whether ChatGPT or Perplexity mention their brand are still doing it by hand: typing the same three prompts into four chat windows every so often and eyeballing the results. That produces no trend line and nothing a client will accept as a report. GeoLook is an open source GEO tool built to replace that habit with something measured. It samples real answers across a set of AI engines on a schedule, tracks whether a brand shows up, and turns whatever's missing into a concrete plan instead of a vague sense that something should be done.

GEO here stands for generative engine optimization, meaning the work of getting AI answer engines like ChatGPT, Perplexity, and Gemini to mention and cite a brand when someone asks a category question, not a geography tool and not classic search engine ranking.

What GeoLook actually does

The repository, maintained by a developer going by aigclink, runs entirely on a local machine. It is a Python script and a small dashboard, not a hosted product. Point it at a website (or a brief document if there's no site yet) and it walks through four stages: status, diagnosis, strategy, and verification.

Status pulls real sampled answers from up to 17 engines, ten of them automated through API keys and seven checked manually, including Google AI Overviews and the Chinese search assistant Metaso. It reports mention rate, ranking, and citation share, and lets a reviewer browse the raw sampled answers rather than trust a single aggregate score. Diagnosis runs a site audit organized as a four-layer dependency chain: whether AI crawlers can access the site at all, whether the site's structure orients them, whether the content is actually understandable to a model, and whether individual passages are quotable on their own. A site that fails the access layer makes everything downstream invisible, so GeoLook computes the fix order rather than leaving that judgment call to the reader.

Strategy converts the diagnosis into tickets with acceptance criteria, a risk grade, and a deep link to the specific question each fix is meant to move. Verification re-crawls the site and re-samples engines to check whether a ticket actually worked, and reopens it automatically if a metric regresses. That loop, from a raw sample to a reopened ticket, is the part most GEO tools skip.

Free, self-hosted, and where the data actually lives

GeoLook is MIT-licensed and runs as a script on the user's own machine, with a lightweight dashboard served locally. There's no account creation, no vendor cloud, and no subscription. Every sampled answer, every ticket, and every audit result lands in a local work/ folder as plain JSON and Markdown files. The README is explicit that git init is the backup strategy, since nothing leaves the machine unless the user chooses to publish it.

This is also where it splits clearly from AIOS Guide's earlier writeup on AI Search Console, a paid, hosted GEO tool from the team behind search-console.ai. AI Search Console runs the sampling on its own infrastructure, on a schedule, and hands back client-ready PDF reports through a metered credit system starting at $45 a month. That's the right shape for an agency managing several brand accounts who wants a polished report with no setup. GeoLook is the opposite trade: free, but it runs on hardware the user maintains, on an API budget the user pays for directly (which can be zero, since manual sampling works with no keys at all), and any client deliverable has to be assembled rather than downloaded. Anyone choosing between the two is really choosing between paying for convenience and paying with setup time.

Traction and verification

As of today, the repository sits at 466 stars and 114 forks on GitHub, with 22.5 megabytes of source, active issue tracking, and a pushed commit as recently as three days ago. It carries an MIT license and lists requests, beautifulsoup4, and lxml as its only three third-party dependencies. The star count has moved fast since the project's late-July appearance; two earlier scan snapshots recorded 151 and then 283 stars within roughly a day of each other. Star counts are a proxy for developer interest, not for whether the tool's citation numbers hold up in production, but the growth rate is real and worth noting for anyone deciding whether to invest setup time now versus waiting.

Requirements and setup

GeoLook needs macOS or Linux (Windows works through WSL, since the code relies on fcntl file locks) and Python 3.9 or newer. Installation is pip3 install requests beautifulsoup4 lxml followed by python3 scripts/geo.py ui, which opens a local dashboard at 127.0.0.1:8765. Crawling, auditing, ticket generation, and asset creation all work with zero API keys configured; adding one key for a Chinese-market engine like DeepSeek or GLM unlocks automatic question-bank generation and AI-assisted first drafts. For engines with no public API, a Chrome extension handles manual sampling by loading a queue of buyer-intent questions and extracting the finished answer with its citations in one click.

Who this is genuinely for

GeoLook fits a technical marketer, developer, or small agency comfortable running a Python script and reading through a dashboard, who wants full ownership of GEO data and doesn't mind doing their own reporting. Its coverage of Chinese-market engines (Baidu, Doubao, DeepSeek, Kimi, MiniMax) alongside the usual Western set is a genuine differentiator for any brand selling into both markets, something the paid SaaS options in this category tend to skip.

It is not for someone who wants a client-ready report with zero setup, or a non-technical operator without time to run through a CLI-driven onboarding wizard. It's also, by its own README's admission, a single-machine tool with no team accounts or collaboration layer, so a larger in-house team managing several brands will hit friction the hosted alternative doesn't have. And like any sampling-based GEO measurement, a single answer from a single engine is noise; the tool's own documentation is upfront that mention rate and citation share only mean something once aggregated across dozens of questions and repeated rounds, not read off one lucky or unlucky prompt.