See what ChatGPT & Gemini
actually cite.

It's a free AI citation checker — a Chrome extension that passively records what ChatGPT and Gemini do behind an answer — the hidden search queries they run before responding, every source they read, which of those sources they credit by name versus quietly use without credit, and every brand, product, or local business they mention. You can look at any single answer in detail, or track your brand against named competitors across a fixed set of prompts on a repeatable cadence for ongoing GEO (generative engine optimization) visibility. Everything is stored locally in your browser.

An AI citation checker — sometimes called a citation finder — is a tool that shows you which sources a large language model like ChatGPT or Gemini actually read and credited while generating an answer, instead of leaving you to guess. CitoSkeleton is a free citation checker built specifically for this: it passively captures the hidden fan-out queries, the sources read, and the sources credited behind every AI answer, entirely inside your own browser, then rolls that up into brand visibility and share-of-voice tracking against named competitors.

Free, no account, no login, no subscription. There's no server for an account to exist on — everything runs and stays inside your browser.

ChatGPT works fully today: fan-out queries, cited and fetched sources, products, local businesses, brand detection, the complete answer text. Gemini works for everything except fan-out queries — Gemini's web app simply doesn't expose them, which is a limitation of Gemini itself, not a gap in CitoSkeleton. Perplexity, Claude, Grok, and Google AI Mode are on the roadmap and shown as "coming soon" where they appear — they are not functional yet, and CitoSkeleton doesn't pretend otherwise.

No. CitoSkeleton makes no network requests of its own — there's nothing to send data to, because there's no server behind it. It only reads network responses your browser already received because you were using ChatGPT or Gemini normally, and stores what it reads in your browser's local database (IndexedDB).

One question you type into ChatGPT often becomes several hidden searches behind the scenes — a prompt like "best trading app in India" can trigger 5 or more sub-queries and pull in 15+ sources before you see a single word of the answer. Those sub-queries are the fan-out. Until you can see them, you have no idea what the model actually went looking for on your behalf.

Cited means the model named that source directly in its answer. Fetched means the model read the source but never credited it. Mentioned means it's referenced without a live citation. The gap between fetched and cited is usually the single most useful thing in the tool — it's a site the model trusted enough to read, but not enough to name.

A content script running in the page's own JavaScript context patches fetch and XMLHttpRequest before ChatGPT or Gemini's own code runs, then copies (tees) the network response as it streams past. It doesn't modify, block, delay, or interfere with your conversation in any way — it's a read-only copy of traffic your browser was going to receive regardless of whether CitoSkeleton was installed.

The raw, unmodified network response is compressed and stored first, as an immutable record — before anything tries to parse it. A separate adapter then turns that raw data into the structured view you actually use (queries, sources, brands, products). If a format change breaks the adapter, nothing you've already captured is lost: the raw records are still sitting there untouched, and a "Reprocess all" pass re-derives every one of them the moment the adapter is fixed — with zero re-browsing required.

A Campaign is a fixed, named measurement setup: your brand and URL, up to several named competitors and their URLs, a fixed set of prompts, and which engines to run them on. Once you lock a campaign, that definition stays fixed so results are comparable day over day — editing it afterward is treated as a "this breaks your trend line" action that CitoSkeleton warns you about, not something it lets you do silently. It's built for repeatable tracking, not one-off curiosity.

Visibility is the percentage of responses that mentioned your brand at all — responses mentioning the brand divided by total responses. Share of Voice is your brand's mentions as a percentage of every tracked brand's combined mentions — your mentions divided by everyone's mentions. Both are calculated using the same published formulas as Peec.ai, a well-known paid GEO analytics tool, so the numbers are apples-to-apples comparable with anyone using that category of product.

Because a trustworthy 0–100 sentiment score needs either a second-pass AI judge or a hand-built lexicon, and neither is in place yet. Shipping a fabricated number in a measurement tool would be worse than admitting the gap, so CitoSkeleton simply doesn't measure it. This is a deliberate quality stance, not an oversight — the same standard that keeps Position honest (see below) rather than showing every metric as "complete."

Yes. You can type prompts one per line, or import a CSV or Excel (.xlsx) file directly. Columns are matched by header name — Prompt, Query, Tags, Industry, Category, and similar — so a real prompt sheet from a research team usually imports with zero reformatting.

Campaign runs submit prompts through your real, logged-in account, at a human-like pace with randomized delays rather than a burst — and CitoSkeleton is upfront that this is a real automated action with real account-safety implications, not something to treat casually. Before a run starts, it shows you exactly how many submissions it will make (prompts × engines) and warns you if that number is large — 60 or more. Nothing ever fires unattended; every run needs a manual click, and the most CitoSkeleton does on its own is remind you once a day if a scheduled run hasn't happened yet.

Several formats depending on what you need. A full Excel workbook with 11 separate worksheets covering everything captured — summary, audit log, source analysis, unique domains, the complete source list, every fan-out query, every brand mention, every product, every local business, every entity, and the full answer text of every response. A clean HTML document or a print-ready PDF for single conversations or a bulk-selected set, with ChatGPT's inline citations rendered as real numbered footnotes and product images embedded where shown. Detailed CSV and raw bulk JSON. And every drill-down table in the dashboard has its own one-click CSV export.

Because Chrome Web Store rules prohibit remotely-hosted executable code — CitoSkeleton can't push a server-side fix when OpenAI or Google changes their response format. If only parsed data were kept, a format change would silently corrupt or lose new captures until a new store submission passed review, which can take days. Keeping the raw payload immutable means the parser can be wrong today and correct tomorrow without losing a single capture in between — "Reprocess all" re-derives everything from data that was never touched by the broken parser in the first place.

Structurally, not by a filter you can accidentally leave off. A response captured as part of a campaign run carries that campaign's identity as part of its stored record; a response from ordinary browsing doesn't. That distinction is enforced at the point of storage, so a campaign's numbers can never be quietly inflated by an unrelated question you happened to ask that day, and casual browsing can never contaminate a tracked metric. It's engineered in, not bolted on as an afterthought.

Honestly, rather than with a fake rank. Position is the average rank by order of first appearance, averaged only across the responses where the brand actually showed up. A brand that never appears isn't assigned a last-place rank to fill in a number — it's excluded from that average and the gap shows up correctly in a lower Visibility score instead. Inventing a rank for a brand that was never mentioned would misrepresent what actually happened.

Yes — they're hardened against formula injection. A CSV cell that starts with =, +, -, or @ can be interpreted by Excel as a live formula rather than plain text, which is a genuine, under-discussed way exported data can be turned against whoever opens it. CitoSkeleton's exports are built to prevent that.

Network access is scoped to three domains and nothing else: chatgpt.com, chat.openai.com, and gemini.google.com. It has no broader host permissions, makes no requests of its own, and there's no account or cloud infrastructure for it to phone home to. Compare that to the paid GEO tools in this category — Peec.ai, Semrush AIO, Profound, and similar — which require sending your prompts and competitive strategy to their servers and charge a recurring fee for the privilege. CitoSkeleton has no server to send anything to.

No — storage is automatically capped, defaulting to the most recent 2,000 captures, and the cap is configurable if you want to keep more or fewer. There's also a one-click "clear everything" if you want to start over.

Stated plainly rather than left ambiguous: Perplexity, Claude, Grok, and Google AI Mode are visible on the roadmap but not functional — Google AI Mode in particular is a distinct surface from regular Google or Gemini search that does expose real fan-out data, making it the next major target. Sentiment scoring isn't measured, for the reasons above. Visual trend charts over time aren't drawn yet either, though the underlying data for them already exists in every capture. Anything shown as "coming soon" in the product is exactly that — not a working feature with a label slapped on it.