machinesloveyou runs the questions your buyers actually ask, across seven answer surfaces, and keeps the whole answer as evidence. Then it reads your CDN to see which crawlers came back. Self-hosted, on your own hardware, on sources that cost nothing.
What's the best buy-here-pay-here dealer software?
The same prompt returns a different shortlist in Miami than it does in Berlin. machinesloveyou detects the language your market actually reads and runs each surface where your buyers are, not where your office is.
A model does not visit your website before it answers. It repeats what it already absorbed, or it fetches three sources and quotes them. Both paths are measurable. Neither one appears in a rank tracker.
For each engine machinesloveyou records whether you were mentioned, whether a source URL was attached, where you landed in the ordering, and what was kept as proof.
| Engine | Mention | Cited source | Rank in answer | Evidence kept | Run mode |
|---|---|---|---|---|---|
| ChatGPT | Tracked | URL captured | Ordinal | Full answer | Search + memory |
| Claude | Tracked | URL captured | Ordinal | Full answer | Search + memory |
| Gemini | Tracked | URL captured | Ordinal | Full answer | Search + memory |
| Perplexity | Tracked | URL captured | Ordinal | Full answer | Live search |
| Google AI Overviews | Tracked | URL captured | Ordinal | Snapshot | SERP capture |
| Brave Answers | Tracked | URL captured | Ordinal | Snapshot | Independent index |
| DuckDuckGo | Tracked | Result set | Position | Snapshot | Independent index |
| Microsoft Copilot | On the roadmap | — | — | — | Not shipped |
Answer runs are cheap to fake and expensive to trust. Three things separate a measurement from a screenshot.
A model reciting training data is not a citation. Each run records which mode produced the answer, so trend lines never mix the two.
A second model checks the claimed mention against the stored answer. Hallucinated wins are recorded as misses, which is the only way the number stays worth reading.
Full text, sources and timestamp land in your Postgres. When a rival appears in month three you can open month one and see exactly what changed.
AI crawlers never run JavaScript, so no analytics tag will ever see them. machinesloveyou reads the edge directly and tells you which operator came, how often, and how deep.
| Operator | User agent | robots.txt | Reads / 7d | Deepest path |
|---|---|---|---|---|
| OpenAI | GPTBot | Allowed | 1,284 | /pricing |
| OpenAI | OAI-SearchBot | Allowed | 612 | /compare |
| Anthropic | ClaudeBot | Allowed | 486 | /docs |
| Perplexity | PerplexityBot | Allowed | 377 | /blog |
| Google-Extended | Blocked | 0 | — | |
| Meta | Meta-ExternalAgent | Allowed | 205 | / |
| Apple | Applebot-Extended | Allowed | 94 | /features |
| Amazon | Amazonbot | Allowed | 58 | / |
| ByteDance | Bytespider | Blocked | 0 | — |
| Common Crawl | CCBot | Allowed | 41 | /about |
Demand, answers, condition and work. Each module hands the next one its evidence, so nothing in the queue arrives without a reason attached.
Where demand exists and how far you are from it.
Whether the machines say your name, and whose name they say instead.
The condition of the thing the machines are reading.
The queue that turns all of it into work someone can ship.
Every number in machinesloveyou traces back to a source you authorised yourself. There is no scraped panel, no modelled estimate standing in for a measurement.
| Source | What it gives you | Refresh | Cost |
|---|---|---|---|
| Google Search Console | Queries, pages, positions, CTR — Google's own numbers | Daily | Free |
| GSC URL Inspection | Google's indexing verdict for a specific URL | On demand | Free |
| Google Analytics 4 | Sessions, engagement and revenue per landing page | Daily | Free |
| PageSpeed Insights | Lab performance and Core Web Vitals per URL | On demand | Free |
| Chrome UX Report | Six months of field data — yours and your rivals' | Monthly | Free |
| Google Ads API | Historical search volume and competition per keyword | Monthly | Free |
| Google Knowledge Graph | Whether Google recognises you as an entity | Weekly | Free |
| Wikidata SPARQL | Entity existence matched by domain, not by name | Weekly | Free |
| Bing Webmaster | The queries and backlinks Microsoft sees | Daily | Free |
| Brave Search | An index that is not Google's, for cross-checking | On demand | Free |
| Common Crawl graph | Backlinks extracted from the open web archive | Monthly | Free |
| Open PageRank | Authority proxy for every linking domain | Monthly | Free |
| Cloudflare Analytics | Which AI crawlers hit you, and which pages they read | Daily | Free |
| Microsoft Clarity | What humans do once an answer sends them to you | Daily | Free |
| Hacker News / Algolia | Brand mentions that never became a link | Weekly | Free |
| Your own crawler | Render audit, indexability signals, robots policy | On demand | Free |
| OpenRouter | Model routing for answer-engine runs | On demand | Metered |
Sixteen of the seventeen sources above are free at any volume you will realistically use. The only metered line is the model calls you choose to make.
Demand on the left, proof on the right. Every panel is fed by a source you authorised yourself, and every claim keeps the record that produced it.
No untraceable score. Every number opens onto the run that produced it.
Ranked by measured impact, not by how loud the finding felt.
The loop the instrument runs. Each step hands the next one something it can act on, and the last one closes only when the measurement agrees.
Pull the demand and the answers.
Pick one to see what it holds. Each panel is fed by a source you authorised yourself, and every number opens onto the run that produced it.
Repeatable runs across seven surfaces, each answer stored whole so the claim and the receipt live together.
7 surfaces measuredThe whole instrument is exposed over MCP. An agent can read any dataset, queue any job and write the report — always inside one project, never across them.
$seo_projects2 projects · multi-tenant, every query scoped by slug
$seo_data --set radar38 datasets exposed · answers, citations, entities, crawls
$seo_run_job url-inspection34 jobs · queued through pg-boss, no cron required
$seo_report --weeklyreport written to disk and to the database
$
Compared by capability rather than by vendor, because the useful question is not which tool is better — it is which half of the picture you are currently missing.
| Capability | Rank trackers | AI visibility tools | machinesloveyou |
|---|---|---|---|
| Classic search positions | Yes | Rarely | Yes |
| Mentions inside AI answers | No | Yes | Yes |
| The source URL the model cited | No | Sometimes | Yes |
| Raw answer stored as evidence | No | Rarely | Yes |
| Memory vs live search, labelled | No | No | Yes |
| Entity presence in Wikidata and Google | No | No | Yes |
| AI crawler reads from your own CDN | No | No | Yes |
| Search Console joined to revenue by page | Partly | No | Yes |
| Drivable by an agent over MCP | No | No | Yes |
| Runs on your hardware, your database | No | No | Yes |
| Priced per seat | Yes | Yes | No seats |
Postgres, a queue and a worker in Docker Compose. A small VPS is enough to hold every project you will run this year.
You authorise each source with your own accounts. Evidence lands in your database and stays there — nothing is copied to a vendor.
No per-seat licence and no credit meter on measurement. The only variable cost is the model calls you choose to make.
On one machine you already own. Postgres, a job queue and a worker in Docker. There is no control plane, no tenant of ours holding your data, and nothing phones home.
The measurement sources are free APIs you authorise with your own accounts. The only variable cost is the model calls you decide to make when running answer engines, billed by whichever provider you route to.
No. Crawler reads come from a CDN adapter, and Cloudflare is simply the first one shipped. Fastly, CloudFront or plain access logs feed the same table through the same interface.
That is the point of the MCP layer. Thirty-eight datasets and thirty-four jobs are exposed as tools, every one of them scoped to a single project, so an agent can read the evidence and start the work.
Every run is labelled memory or live search, and a separate judging pass checks the mention against the stored answer. A model reciting its training data is recorded as exactly that, not as a citation.
Connect Search Console, point it at your CDN, and run the first sweep. The instrument tells you the truth on day one, which is usually uncomfortable and always useful.
Analyze your site