Framing analysis — how AI talks about you
Being named is not the same as being recommended. Framing analysis classifies every mention in every answer — were you the pick, a neutral list entry, or named with a caveat? — and charts the mix per brand as a stacked bar, with the verbatim phrases AI keeps using for each of you underneath. It's sentiment analysis with the invented precision removed.
How it works
Every mention, classified
A judge model reads each answer and tags every brand mention: recommended pick, neutral listing, or named-with-caveat. Discrete events, counted.
Charted as what it is
Per brand, a stacked bar: length is total mentions, segments are how those mentions were framed. You against every tracked rival in one view.
The words themselves, kept
Under each bar, the verbatim phrases from the answers — “cheapest entry point”, “built for larger accounts”, “stronger prediction features”. The framing engines repeat is the framing buyers hear.
In depth
Why counts beat a mood score
A “sentiment: 65” has no units, no interval and no way to be audited. “7 picks, 11 neutral listings, 2 caveats” is checkable — every count links back to answers you can read in the inspector — and it tells you what to fix: caveats name the objection AI keeps repeating.
Caveats are the actionable part
When engines keep naming you with the same reservation — pricing, a missing integration, an old limitation you've since fixed — that's a page fix, not a mystery. The verbatim phrase tells you exactly which sentence your site needs to correct in public.
Framing shifts show early
A rival moving from neutral listings to recommended picks often precedes their rate climbing. Because the classification re-runs weekly, framing movement shows up while it's still cheap to respond to.
What we won't print: a single sentiment number per brand. At 60–150 answers, a 0–100 score has error bars wider than most of the differences it claims to show — dressing that up as a metric would break the one promise this product runs on.
Common questions
Does CiteGraph do sentiment analysis?+
Yes — as framing analysis. Every mention is classified as a recommended pick, a neutral listing, or named-with-a-caveat, charted as stacked bars per brand, with the verbatim descriptive phrases underneath. What we deliberately don't produce is a 0–100 sentiment score, because at real sample sizes that number is noise presented as precision.
What does a caveat mention mean?+
The answer named you but attached a reservation — “pricier than X”, “lacks Y”. Caveats are the most actionable finding in the panel: they quote the objection AI repeats to your buyers, which is usually a sentence your own site can correct.
Can I verify the classification?+
Every count is auditable — the Answers inspector stores each answer verbatim with your brand highlighted, so you can read the mentions behind any bar yourself.
Which plans include framing analysis?+
Every plan. It runs in the same extraction call as competitor discovery, on every scan.
See it on your own domain — first dossier in about a minute.