Skip to content

Journal

What the numbers are actually of.

Twenty-three essays on crawlers that never run JavaScript, people who arrive from an assistant, hashes that expire at midnight, and agents that read rollups — not rows.

A dark server corridor with amber status lights receding into shadow

·6 min

GPTBot never runs your analytics script

Training and search crawlers fetch HTML and leave. If your only collector is a browser snippet, those fetches never existed — and most privacy-first tools still miss them.

AI crawlersMiddleware

Hands on a laptop at night, lit only by the screen

·5 min

ChatGPT referrals are not Direct

People who click a citation inside an assistant often arrive with a stripped referrer. The rest show up as chatgpt.com, perplexity.ai, or utm_medium=ai_search. Filing either under Direct hides a channel.

human_via_aiAttribution

A legal document and a thinner stapled packet on a black desk

·5 min

Make “private” a document, not a slogan

Every Vector 5 site publishes a machine-readable collection manifest: what is stored, how uniques are counted, and which questions the system cannot answer. Auditors and agents can fetch it.

PrivacyManifest

A compact home workstation at night with a single monitor

·5 min

Ollama insights see aggregates. Never a visitor.

Weekly write-ups run on rollups — counts, ratios, deltas. Point the provider at Ollama and nothing leaves the machine. Event-level rows never enter the prompt.

OllamaInsights

Two computer terminals facing each other in a dark room

·6 min

An MCP server so agents read the same numbers you do

Six read-only tools — stats, pages, sources, AI traffic, goals, insights. Claude, Cursor, and Codex can brief you from observed rollups, not from a CSV you pasted.

MCPAgents

Close-up of a black network switch and a faintly glowing cable

·5 min

Count the visitors who never execute JavaScript

Crawlers, agents, and no-JS browsers share one property: your script did not run. A small middleware in the request path is how they still appear in the data.

MiddlewareNo-JS

A black analog clock at midnight beside a small pile of white salt

·4 min

A visitor hash that forgets you at midnight

Unique visitors without cookies means a hash, not an ID. Vector 5 salts per site, rotates at 00:00 UTC, and drops the IP. Tomorrow you are new — even to us.

PrivacyUniques

A dark library aisle with one open book in a shaft of light

·6 min

Crawl-to-referral is the only AI SEO number that is not a guess

Models fetch pages. Sometimes those pages later send you a person. The ratio is observable. Rank-tracking for ChatGPT is not.

AI SEOCrawlers

A blurred night crowd on a wet street, no face recognizable

·5 min

Why we do not fingerprint “just a little”

A canvas hash, an audio context, a font list — each is sold as a cookieless unique. Together they are a tracker that outlives the cookie banner you avoided.

PrivacyFingerprinting

A dark studio with a large monitor showing an almost empty chart

·4 min

The empty dashboard is the product

Vector 5 never seeds sample traffic. Until a real pageview or crawler hit arrives, the blanks stay blank. That is the privacy model showing its work.

ProductTrust

A human hand and a robot hand reaching toward the same cup

·5 min

Bounce rate lies when agents shop for you

Playwright adding a pair of shoes is not a bounce. It is also not a conversion you should mix with Safari. Actor filters keep both numbers honest.

ActorsGoals

Night shipping containers, one door open onto warm light

·4 min

The install is still docker compose up

Postgres, the Go ingest, the dashboard. One file. If the product needs a salesperson to turn on, the privacy claims are already downstream of procurement.

Self-hostDocker

A small home server on a shelf beside books and a plant

·5 min

Self-host so the visitor never becomes inventory

The same code runs on your laptop and in our cloud. The manifest does not change. That is the point of a clean-room server you can read.

Self-hostPrivacy

A printed page almost entirely covered in black redaction bars

·5 min

What Vector 5 cannot know — and why that is the feature

Who someone is. Whether they came back after midnight UTC. A city, unless you turn it on. Anything about people who block the request. The dashboard prints this next to the numbers.

PrivacyLimits

A cream printed page with one gold-highlighted sentence

·6 min

PerplexityBot, OAI-SearchBot, and the citation graph on your origin

Search-shaped crawlers fetch so an assistant can cite. Treat them as a graph of pages the models have actually read — not as “bot noise” to exclude and forget.

AI crawlersPerplexity

An empty institutional hallway with identical doors and fluorescent lights

·5 min

Direct traffic is a measurement failure

Bookmarks exist. So do stripped referrers, assistant clicks, and in-app browsers. Calling the residue Direct is how a channel disappears from the meeting.

AttributionDirect

Two antique keys on black velvet, one gold-toned and one silver

·5 min

AGPL server, MIT edges: why the licence splits

The ingest and dashboard are AGPL-3.0 so improvements flow back. The tracker, middleware, MCP, and skills are MIT so embedding them never licenses your product.

Open sourceLicence

Five ceramic spheres in a line: white, blue, amber, green, gray

·4 min

One dropdown. Every report.

human, human_via_ai, ai_crawler, ai_agent, bot, or all. Visitors, pages, sources, countries, devices, goals — the same filter, everywhere.

ActorsProduct

An unlabeled topographic map and a closed brass compass

·4 min

Country without city: geo that does not become an address

The IP is used for a country — and optionally a region — then discarded. City is off by default. Precision is not the same as insight.

PrivacyGeo

An open suitcase packed with folders and a hard drive in a doorway

·4 min

Import. Export. Leave.

Bring history in as CSV. Take every aggregate out as JSON. The schema lives in the repository. Lock-in is a product failure we refuse to ship.

PortabilityData

A linen notebook, mechanical pencil, and keyboard in morning light

·5 min

Skills for Claude, Cursor, and Codex

MIT prompts that sit next to the MCP server: audit crawler access, find cited-but-not-clicked pages, write the weekly briefing from observed rollups only.

MCPSkills

An abstract close-up of a dark search field with a soft blue caret

·4 min

utm_medium=ai_search is a real channel

When an assistant bothers to tag the click, believe it. Vector 5 treats ai_search and ai-assistant as first-class media, not campaign junk.

UTMAttribution

A headless mannequin in profile facing a camera in a dark studio

·5 min

Headless agents are not crawlers

GPTBot reads. Playwright acts. Mixing them into “bots” hides the thing that is already walking your checkout.

AgentsActors