CiteSight

How GEO audits work

Why GEO

Search is no longer just ten blue links. People ask ChatGPT, Claude, Gemini, Perplexity, and AI overviews “who is X”, “what is X”, “notable Y people or orgs”, and “X vs Z”. Generative engine optimization (GEO) is the discipline of becoming the entity those systems mention — accurately, positively, and with citations. CiteSight is the audit layer for any public reputation entity: a company, a person, a nonprofit, a creator. The job is controlling AI/digital presence — appear more where the record is accurate, correct what models get wrong, or fade leftover associations at the source. Measure the gap, pick a goal, prescribe the work, then watch it on a cadence.

Probes

Each run executes a small, type-aware set of prompts against every registered provider: identity, field/category shortlist, peer comparison, trust or credibility, and alternatives. A person gets “Who is X?” and “alternatives to hiring or following X”; an NGO gets mission, raters, and impact; a business keeps the commercial probes. Live mode calls the vendor API when a key is present. Otherwise a labeled demo adapter returns realistic sample answers so the product is demoable without credentials. Providers live behind a common interface in src/lib/providers.

Scoring

Overall score = 0.35 × visibility + 0.25 × accuracy + 0.20 × sentiment + 0.20 × share of voice. Weights: visibility 35%, accuracy 25%, sentiment 20%, share of voice 20%.

Visibility averages prominence (primary 100, featured 78, passing 42, absent 0).
Accuracy compares the answer to claimed field/category plus homepage facts and fetched profile pages — especially Wikipedia, LinkedIn, and Crunchbase. If no Wikipedia URL is pasted, CiteSight tries en.wikipedia.org/wiki/Name and Name_(company) and collects matching articles. LinkedIn company (or person) URLs from the setup form, homepage links, or schema sameAs are recorded even when LinkedIn returns a login wall — CiteSight never invents an About. Optional company-slug guesses are kept only when public HTML actually matches the entity. It does not silently pick a winner when Wikipedia, the claimed website, and AI answers could be different entities (Lovable the vibe-coding company vs Lovable lingerie). The report asks Which entity did you mean? and Correct / Appear then run against the one you choose. If a single Wikipedia article matches the site domain and models agree, the ask is skipped.
Sentiment is a keyword window around the mention — directional, not a courtroom.
Share of voice is mention-bearing answers versus named peer hits.

Monitoring

Start Monitor from any report ($49/mo, two clicks). Demo checkout activates a weekly watch without a card; if STRIPE_SECRET_KEY is set, the same button uses Stripe. Due watches re-run the same probes via /api/monitoring/tick or immediately with Run now on /monitoring. Each completed run stores scores, a “what changed” diff, and in-app alerts on score drops, a peer newly appearing, a description that flipped, or a new negative / inaccurate description.

Trust & limits

AI answers are under-regulated and gameable: the sources a model can retrieve or remember shape what it says. CiteSight exists so an entity can audit that surface, legitimately appear / correct / fade, and defend against false or hostile narratives.

We optimize real public presence. We do not fake consensus, and we do not plant false claims about third parties. Assisted drafts tell you to update pages you control, request factual corrections, and cite independent sources. Every probe card shows source evidence (or the lack of it) and a “why the model might say this” note so the work stays explainable.

Limits: models can still recite archives and training data. Demo adapters are labeled. Wikipedia edits stay human-in-the-loop. Billing and outbound email are stubs.

Presence goals

CiteSight is not only “get cited more.” Each entity has a presence goal that biases the work queue and assisted drafts:

Appear / grow — increase accurate mentions and share of voice (listicles, comparison pages, schema, profiles).
Correct — fix wrong, outdated, or harmful descriptions (canonical bio/About, corrections page, source refresh, raters).
Fade / reduce — reduce unwanted prominence where source-level work can help. Models cannot delete the internet or training data. Fade actions update pages you control, prefer newer authoritative sources, and compete with leftover associations. Old articles can still be recited.

Set the goal on setup or on the report. Changing it re-ranks the latest run’s recommendations without requiring a new scan.

Work plan & assisted drafts

Each run upserts a prioritized work queue from the recommendation catalog. Business audits get schema, About/FAQ, NAP, profiles, Wikipedia/Wikidata, LinkedIn company or profile alignment, listicles, citation assets, comparison pages, reviews, and llms.txt. Person and creator audits add a consistent bio, Person JSON-LD, interview pages, speaking/booking pages, and knowledge-panel / Wikidata work. Nonprofit audits add impact reports and charity-rater profiles. Tasks keep status (todo / in progress / done), can be accepted or skipped, and in Modular mode can be marked for à la carte funded execution. Managed mode auto-includes P0/P1. Generate copy-ready drafts either way. Register a publisher in src/lib/actions.ts when you want to auto-apply them.

Prompt Lab

/lab is an optional A/B workbench on the same provider registry — not the core product. It does not write audits, watches, or work-queue tasks. Comparison is mechanical (word count, sentence overlap, a stub rubric).

Packaging experiment

CiteSight can sell as Modular (monitor subscription + à la carte actions) or Managed (one bundle: monitoring plus every P0/P1 action). Same work queue; PACKAGING_MODE or a per-entity toggle switches the plan UI. Checkout is stubbed so both conversion paths are demoable without a payment processor.