The Classifier
How GEOforge categorises citation sources so you know whether to reply, pitch, or skip each opportunity.
The Classifier is the agent that decides, for every citation source GEOforge surfaces in CiteForge, what kind of opportunity it is and how you should act on it. You don't run the Classifier directly — it runs automatically when opportunities are discovered or reclassified — but its output drives the badges, filters, and suggested actions you see throughout CiteForge.
What it produces
For each URL, the Classifier emits two labels:
- Opportunity type — the action template. One of:
UGC— a community thread, Q&A, or forum post where you can contribute a reply.Outreach— a page owned by someone else where you'd need to email the publisher or contributor to be included.Other— no useful action surface (your own site, a competitor's site, a profile root, etc.).
- Source category — the kind of source it is. One of:
UGC platformIndustry blogNews / editorialEditorial listicleExpert referenceAssociation / directorySocial profile pageBrand-owned domainCompetitor domainOther
The opportunity type tells you what to do. The source category tells you what kind of thing it is, so you can filter and prioritise.
How it decides
The Classifier tries a fast URL-pattern check first. If the domain matches a known pattern — say reddit.com, medium.com, a .gov site, or your own website — it applies the matching rule and stops there. This is why most opportunities are categorised instantly with no cost.
When the URL doesn't match any known pattern, the Classifier asks an LLM to make the call. The LLM receives:
- The URL and any context snippets from where it was cited.
- Your company's name, website, and industry.
- Your competitor list.
That brand context is what lets the Classifier reliably tag your own site as Brand-owned domain and your listed competitors as Competitor domain, even when their URLs would otherwise look like normal editorial pages.
What the categories mean in practice
| Source category | Typical action |
|---|---|
| UGC platform | Draft a reply with The Pitch Artist and post it from a real account. |
| Industry blog | Outreach — pitch a guest post or expert quote. Author controls publication. |
| News / editorial | Outreach — pitch a journalist for future coverage on the same beat. |
| Editorial listicle | Outreach — ask to be added to the list. |
| Expert reference | Outreach, but slower — these are high-trust references (Investopedia, Britannica, .edu, .gov). |
| Association / directory | Outreach — apply for a member listing if you qualify. |
| Social profile page | Skip. Profile roots aren't actionable; you can't add yourself to someone else's profile. |
| Brand-owned domain | Skip — it's already you. |
| Competitor domain | Skip — you can't earn citations on a competitor's own site. |
| Other | Skip. Wikipedia, GitHub, and similar non-actionable references. |
Why social URLs sometimes split
Social networks get special handling. A LinkedIn company page, an X handle root, or a YouTube channel front page is tagged Social profile page — there's no useful surface to act on. But specific posts (a LinkedIn post URL, an X status, a YouTube video) fall through to the UGC rules and get tagged UGC platform, because those do have a comment surface you can engage with.
Personal LinkedIn profiles (/in/<name>) are intentionally sent to the LLM rather than auto-categorised, so notable experts and journalists can be promoted to Expert reference based on the context where they were cited.
Reclassifying opportunities
If you change your brand profile — for example, you add a competitor, update your website, or refine your industry — the existing classifications won't automatically update. An admin can trigger a full reclassification of every opportunity for the brand using the Reclassify all button at the top of CiteForge → Opportunities.
This is worth doing after any significant brand-profile change, since the Classifier uses that data to apply the brand-self and competitor exclusions.
Tuning the Classifier
Admins can adjust how the Classifier decides borderline cases by editing its prompts under Settings → Agent Prompts → Classifier. Two prompts are exposed:
- System prompt — the Classifier's general instructions: what each category means, how to weigh URL patterns against snippet context, and the tone it uses to justify its calls. Edit this when you want to tighten or relax how a category is applied across the board — for example, when you want news-aggregator sites treated as
News / editorialrather thanOther. - User prompt — the per-URL template that wraps the inputs (URL, snippet, brand context) before they go to the LLM. Edit this when you want to change what evidence the Classifier sees — for example, to pass in additional context fields, or to ask it to consider a custom signal.
Changes take effect on the next classification run, not retroactively. After any prompt change, trigger a full Reclassify all from CiteForge → Opportunities — otherwise the existing categorisations will still reflect the old prompt and your overrides will only show up on newly-discovered sources.