Twitter API for Social Listening

The brand lane is the narrow end of the panel, and for reputation it is the one that has to be fast. Watch three things at once: posts that tag your handle through GET /twitter/user/mentions, posts that name you without tagging you through a search on your brand and product names, and posts that misspell you, which matter more than teams expect because an angry customer rarely stops to check the spelling. Exclude your own account from the search with a -from: operator so your own posts do not count as the market talking about you.

Reputation is not a count. A day with twice the usual mentions can be a good launch or a bad outage, and the difference is in who is talking and how. Resolve the authors behind any jump, so one large account and a thousand small ones read differently, and read the replies under the posts driving it, since that is where a complaint either finds agreement or dies out.

A share figure on its own says where you stand; a result is a change in it. To measure a campaign, a launch or a crisis response, take two windows of identical length, a baseline before and a measurement window during, and run the identical query set over both. Because search takes since: and until: bounds, the baseline can be rebuilt after the fact for a launch you forgot to baseline, at the same per-call rate as a live query.

Report three numbers side by side: the change in mention share, the change in engagement share, and the change in reach-weighted share, with the medians beside the means. If all three move the same way, the result is real. If only reach-weighted share moved, one large account probably carried it, which is still a result but a different one. Tying any of this to signups or revenue happens outside this API, by joining your dates against your own analytics, and it is worth doing once to learn which of the three shares tracks your business at all.

Your category queries only find conversations you knew to describe. Trends cover the rest. GET /twitter/trends returns the current ranked trends for one location, about 50 of them, and takes a country name, an ISO code or a numeric WOEID. The full list of supported locations and their WOEIDs comes from GET /twitter/trends/locations. Pass a category instead of a location to read one of the Explore tabs, such as news or sports, or to keep only the trends X labels business and finance.

Snapshot the markets you sell into on a schedule and keep every snapshot, because the useful signal is movement: a term entering the list in one country before another, or a term near your category climbing for three snapshots in a row. Ten markets read hourly is 7,200 calls a month, about $5.76. When a trend touches your category, hand the term to the search lane and read the posts behind it. The trends API guide builds trending-topic detection by location with code.

The API returns the text of every post, not a sentiment label. That is a choice in your favour: you pick the scoring method, you can rerun last quarter's posts with a better model, and no vendor score changes underneath a chart you already published. Run the scoring as a local pass over the stored text, from a lexicon method for speed to a language model for nuance.

Whatever you pick, measure it before you trust it. Label 200 of your own posts by hand, score the same 200, and count the disagreements. Sarcasm, category jargon and posts that quote a complaint in order to rebut it are where scoring goes wrong on X, and every category has its own share of each. Always publish sentiment beside the volume it was computed on, since a strongly negative reading across nine posts is noise and a mild one across nine thousand is news. The sentiment analysis in Python tutorial walks through a full scoring pipeline, and the sentiment API buyer's guide compares the data sources.

Listening suites ask you to choose a tier before you know how much of the category you need to hear. Here there is no tier. Write the panel down as lanes, give each a query count and a cadence, and the bill follows from the arithmetic in the table above. Start every lane hourly, look at a week of real volume, then tighten only the lanes where an hour of delay would have cost you something. That is usually the brand lane and almost never the category lane.

Two habits keep the number where you planned it. Budget against calls, not tweets, since a narrow query returns fewer than a full page and still bills the call. And check the balance from inside the job with GET /account/me, which is free, so a panel that has drifted above plan raises its own alert before an invoice does. Launch weeks are the one planned exception: raise the cadence for the week, then set it back.

Protected accounts. Posts from private accounts are not public data and appear in no listening product, ours included. On a niche category this can be a meaningful slice of the practitioners talking.

Anything deleted before your poll. A post removed inside your polling interval never enters the corpus. Shorter intervals narrow the gap and never close it, which is worth remembering when the deleted posts are the ones that mattered.

Conversation your query did not describe. This is the largest blind spot by far and the only one you control. A category is described in words you did not think of, so re-read a raw sample monthly and harvest the vocabulary your query is missing rather than trusting a term list written once.

Other networks. This is an X API. If your category argues on several networks, this panel is one input rather than the whole picture, and a share-of-voice number computed here should be labelled as X-only rather than as social.

Impressions. Follower counts are a reach proxy, not a measurement of who saw a post. Weight by them if it helps, and never publish the result as if it were an impression count.