Just telling Claude Code to "look it up" only ever searches the same place — a branching net-surfing method
Introduction
Ask an AI simply to “look it up” and, in most cases, it queries a search engine once, summarises whatever came back at the top, and stops. Plenty of questions are settled that way. But for an investigation where the answer sits somewhere nobody has summarised yet, that approach will almost certainly never reach past the range of a single entrance.
This is a record of the method I established to get past that limit — the branching net-surfing method. Rather than searching once and stopping, you take the proper nouns that appear in the results, feed them back in as the next search terms, and keep extending branches. Establishing this procedure is what let an investigation that had stalled without an answer reach, somewhere in the loop, real data published by a public institution.
First set of queries to the search engine (3-5 from different angles, nothing read yet)
│
├─▶ Branch 1 ── open it and take everything (body, figures, footnotes, links)
│ └─ collect proper nouns (scheme names, institution names, sources for figures)
│ └─ feed those proper nouns back into stage A as new search terms ─┐
│ │
├─▶ Branch 2 (do not skip something as "similar to what I just read") │
│ └─ open another article covering the same topic │
│ └─ find the difference only that article has (a figure, a note) │
│ └─ collect proper nouns ───────────────────────────────────┤
│ │
└─▶ Branch 3 … │
▼
back to stage A (one lap, and the branches multiply)
│
the answer to the stalled question turns up somewhere in this loop
Stages A, B and C — spread branches, choose branches, take everything
The first thing to do is not to settle for a single query, but to throw several queries from different angles at the search engine. The phrasing practitioners use, the academic phrasing, phrasing that asks about the mechanism itself, phrasing that includes figures and units — at least three to five variants, collecting only titles, URLs and snippets as a list, without reading any of the contents yet.
Next, choose what to open from that list. The judgement most easily got wrong here is skipping another article on the same topic because “it looks like more of the same”. In practice, several articles covering the same topic often carry different figures, different periods, and different notes. In one case involving a specialised term, among several similar explanatory articles on the same topic, only one actually mentioned that term concretely. Had I read only the others, that branch would never have existed.
From a page you open, take everything, not just the body text. Capture figures and graphs as images and read them. Notes and legends on a figure frequently carry a source or a period that the body text never states. The list of links is itself the set of branches to open next.
Stage D — the body of the loop
In one phrase, the heart of this method is net-surfing. From information that looks related, you read off further useful information, and search again on that term. Rather than digging in one place forever, you branch outward and let the material expand. The only difference is that I imposed this repetition on myself as a procedure, instead of leaving it to whim.
The proper nouns picked up from the pages opened in stage C — the name of a scheme, the name of a statistic, an author’s name, the name of an institution cited as a source — go straight back in as the next search terms. With every lap, the number of branches available to open next actually grows.
There is a case where this repetition paid off. On a market practice in foreign exchange, an investigation into “who actually moves it” had stalled with no answer. Feeding one specialised term back in as a search term (a proper noun for the mechanism by which the market’s rate is set) led to trader-level transaction data published by a regulator, and the question was settled with real data. Looking only within the range of the first set of search results, that path would never have been reached.
Record “not found” separately from “not found yet”
Record not only what you found, but also what you failed to reach. Concluding that something does not exist because one search did not turn it up is just writing down the moment you stopped looking as though it were a result. A miss means “not found through that entrance”, not “does not exist”. Change the entrance, and come back to it.
Compare any “rule” you find against a baseline rate before reporting it
Run this method for a while and a moment arrives when it feels as though a rule has come into view. That is the most dangerous point. In one case, while looking at the dates of published events, several happened to fall on public holidays and caught my eye. The hypothesis “perhaps they are aiming at holidays” suggested itself — but recalculating against the proportion of holidays across the whole period, the count that had caught my eye was lower than the baseline rate. Those few cases stood out not because they were holidays but for a different reason (the counts in recent years being disproportionately large).
“I noticed something” and “it really is more or less frequent than the baseline” are two different jobs. Decide the denominator, calculate the rate, and where the counts are small, go as far as writing “indistinguishable” — I have built that extra step into the investigation procedure itself.
One failed method does not mean the investigation is over
Finally, one more stance that holds this method up. When some means of retrieval fails — a direct download is refused, a standard tool cannot read the characters — try another route before concluding it cannot be retrieved. Capture the screen as an image and read it, look for the file in another format, go via a page that serves it openly. Not mistaking the range of one method for the range of what is possible is the precondition for continuing to extend branches.
I have kept records of running this method on other subjects too.
What it takes to add one more branch is not a new tool or advanced search syntax, but the single small effort of writing down one proper noun from the page currently open that can become the next search term. It was that one word that moved a stalled question.