METHODOLOGY
How a trace is built
HOW IT WORKS
1. THE QUESTION IS ASKED PLAINLY
Each AI engine is asked the question with no special instructions. If the engine has a search tool available, it decides for itself whether to use it — we never force a search. This matters: it's what happens when a real person asks, not an artificially forced best case. Some engines answer from memory instead of searching, especially on questions they judge don't need current data.
2. AT LEAST TWO ENGINES MUST HAVE ACTUALLY SEARCHED
A trace is only built if at least two of the engines both produced an answer and confirmed they searched for it. One engine's answer is a single opinion, not something we can call a corpus signal — traces that don't clear this bar are never published.
3. CLAIMS ARE MATCHED TO SOURCES IN TWO STEPS
First, a claim is only considered a match to a source if the thing the claim is about and its specific figure (a price, a number, a named limit) both appear together in the same passage of that source — a lexical gate. Only passages that clear this gate are then ranked against each other by how semantically similar they are to the claim. A source is never credited for a claim just because it's topically related; the gate has to pass first.
CITED means the engine itself declared it read that page. Weight is our own inference, computed afterward from the matching above — a different, weaker kind of evidence than CITED. This is the single most important distinction on this page.
LIMITATIONS — STATED PLAINLY, NOT APOLOGIES
COMMUNITY SOURCES ARE SYSTEMATICALLY UNDER-REPRESENTED
Reddit's robots.txt disallows all bots outright — verified directly, not inferred. Every trace this method produces is built without any Reddit evidence, even on questions where community discussion plausibly shaped what an engine said. This is a known, structural bias in the corpus, not a gap that more crawling would close.
A TRACE RECORDS WHAT AN ENGINE SAID, NOT WHAT IS TRUE
AI engines sometimes assert things that are false. This method does not correct or adjudicate that — deciding what's true is a different job than recording what an engine said, and Source Trace deliberately only does the second one. Every trace page says so directly, not as a footnote.
THE METHOD WORKS ON PRICE AND PRODUCT CLAIMS. IT DOES NOT WORK ON OPINIONS.
This was tested, not assumed. An earlier test batch compared price claims against qualitative ones (e.g. “Webflow needs less maintenance than WordPress”) on the same matching method: 4 of 6 qualitative claims matched zero sources in a 20-source corpus, not because no source agreed, but because independent authors phrase an opinion differently every time, while a dollar figure is nearly always written the same way. A specific number is lexically rare enough to match reliably; a subjective claim usually isn't. Source Trace is scoped to commercial questions with concrete figures for exactly this reason.
THE CURRENT TRACES RAN ON TWO ENGINES, NOT THREE
One of the three engines this method is built to use is temporarily unavailable for budget reasons, not a technical failure. Every trace records exactly how many engines actually searched — never a hardcoded assumption — and will automatically use all three again the moment that changes, with no code change required.