Per-query yield: which searches earn their cost — Open Pulse

At a glance

Published

2026-10-19

Written for

Anyone tuning a listener that feels quiet, noisy or expensive

Category

informative

Reading time

9 minutes

Keywords

per-query yield, listening query cost, search query performance, alert noise, listening quality report

Overview

A listener is not one search. It is a bundle of them: keywords, exclusions, and per-source queries spread across Reddit, Hacker News, X, LinkedIn, Facebook, the job boards, and the review sites. Each query runs on every tick of the schedule, each one costs something to execute, and each one has its own yield. Some pull their weight. Some pull nothing at all.

Almost nobody prices a query. Monitoring tools report totals: mentions captured, alerts sent, leads surfaced. The totals look fine while a handful of the searches behind them return nothing but cost. That is what per-query yield is: the kept signals a single search produces, measured against the runs it consumed. It is the smallest honest unit of a listening setup, and it is where the money goes.

Every search is a bet with a price tag

Retrieval is not free, and this year it got itemised. X moved to pay-per-use API pricing in 2026: a reported $0.005 per post read and $0.010 per user read, with a cap of 3 million post reads per billing cycle on self-serve plans, beyond which you are in Enterprise territory (Superframeworks, October 2026, citing the X developer changelog; a secondary source, so check the current docs.x.com rate card before treating the numbers as exact). Reddit's commercial API is reported at roughly $0.24 per 1,000 calls with a minimum around $12,000 a month, and a single deep thread can consume 15 to 40 distinct requests (Prowlo, October 2026; Reddit publishes no public commercial rate card, so these are reported terms, not an official price list).

This is why, inside Open Pulse, run cadence is the substantive difference between the paid plans. Runs stay cheap where cheap is possible: page shape, off-topic pages, exclusions, blocked domains, out-of-window results, and duplicates are all filtered before any model is called, so filler costs nothing. The expensive step, judging a candidate, only happens to survivors.

The noise problem is at least this bad

Nobody has published a peer-reviewed study of per-query yield in social listening, which is itself worth saying. The closest measured numbers come from a vendor litmus test of Google Alerts: across 148 to 240 tracked companies, roughly 10% of delivered alerts were business-relevant, the rest were passing keyword mentions, stock calls, or duplicates, and about 40% of genuinely relevant business updates were missed entirely. A survey of 230 market-intelligence professionals alongside it found 92% could not rely on Alerts because important updates get missed (Contify, updated around September 2026).

Read those numbers with the right caution. Contify sells a Google Alerts alternative, so the test serves their argument. But the methodology is described: 2,024 alerts manually classified, plus the survey. The figures are the best measured "query yield" numbers in the category, and they are vendor-funded. That is the state of the evidence. Treat any monitoring vendor's precision number with the same scepticism you would bring to a seller's own case study, including ours.

The analogy from a different field lands anyway. In a survey of 100 security operations professionals, 68% said up to three-quarters of their daily alerts were false positives (Workato, October 2026, citing a SOC survey). Incident response is not sales, but the mechanism is the same: a channel that mostly cries wolf trains its audience to ignore it. Lusha put the sales version plainly this year: most buying signals are ignored because triaging them takes longer than reps have (Lusha, August 2026; vendor assertion, not measured data). The alert that fires forty times in a morning gets muted by lunch, and a muted channel is worse than no channel. That argument has its own post.

What yield actually means inside Open Pulse

Yield is not mentions. It is kept signals per search, and it has a visible denominator.

Every run records its step logs: sources searched, hits kept, results skipped as too old or duplicated, signals saved. The quality report turns those logs into per-query yield, how many candidates each search kept and how many were cut, so you can see which of your searches earn their cost and which have never produced a kept signal. The same report, as the API and the MCP server return it, breaks yield down per source too. Next to it sits the excluded list, showing what was dropped and why, in the interface, next to what was kept. When a listener feels quiet, this is how you tell the difference between "the searches returned nothing" and "the filters threw everything away."

A candidate that does not make it is recorded as rejected with its reason rather than stored as a signal: wrong page type, off topic, excluded keyword, blocked domain, outside the window, already seen, ranked too low, scored below the bar, or classified as noise. Rejections from the cheap filters are kept for 14 days. That means a query with zero kept signals still leaves a record, and the record says whether it found nothing or found the wrong things. "Nothing found" is data. Most tools delete it.

Three runs and out

A query that keeps no signals for three runs in a row is disabled automatically, and the disable is reported in the run log and as an in-app notification.

Three design choices sit inside that rule. First, the count is of runs that kept nothing, whether the search came back empty or came back with candidates that never cleared the bar, and a single kept signal resets it. A query that is quiet for one run is unlucky; a query that keeps nothing three runs running is pointed at the wrong shape of conversation, or at a source that is dry for that phrasing. Second, the cutoff is automatic rather than a suggestion in a dashboard, because suggestions to prune dead queries are the easiest advice in software to ignore. Third, it is reported rather than silent, so the quality report shows you the disabled query, and saving the listener's search plan gives every pruned query another chance if you know something the runs do not, like a seasonal market or a launch coming next quarter.

Note what the rule does not do: it never raises your score bar for you, and it never rewrites your queries. Disabling is reversible pruning, not optimisation. Your thumbs do the optimising. Rated signals feed back into that listener's next classification as worked examples, so the definition of relevant becomes yours rather than one a model guessed at on day one.

How to read your own yield

Start with the run log, not the signal list. For a listener that feels quiet, check three things in order: did the searches return anything at all, did the candidates survive the cheap filters, and did anything clear the score bar. A query with candidates and no keeps for weeks is a tuning problem: the vocabulary is close but the intent is not, so adjust keywords or exclusions. A query with no candidates is a sourcing problem: the source is dry for that phrasing, or the window is too tight. A query disabled after three runs did the work of telling you it was done; thank it and replace it.

The per-source breakdown answers a different question. Recall the source labels in the listener wizard: Reddit and Hacker News hand over full text, LinkedIn, X, and the job boards give snippets, Facebook is best effort. A snippet source will always yield worse than a full-text source for intent judgement, and that is not a bug in the source or in your plan. Snippet versus full text is its own subject. Yield measured per source keeps you from comparing them unfairly and deleting a source that was never going to behave like the others.

What the industry shows instead, and why per-query is rare

The category's ROI literature is all program-level. The expanded formula going around this year runs (Direct Revenue + Cost Savings + Risk Mitigation Value + Competitive Advantage Value minus Total Investment) over Total Investment (Syncly, March 2026), and the 2026 guides attach analyst-flavoured numbers to each stream (Digiflute, 2026; the cited Sprinklr, HubSpot, Forrester, and Gartner figures are unverified, so we do not quote them here as facts). Brand24's own 2026 cost guide prices the category in tiers, free tools at $0, SMB at $40 to $300 a month, mid-market at $300 to $1,500, enterprise at $2,000 plus, and names the two cost drivers honestly: mention caps that trigger extra charges during volume spikes, and tracked keyword or project counts as the tier driver (Brand24, updated April 2026).

Program-level math is real math, but it hides the query. A setup can have positive ROI on the whole while three of its eleven searches burn budget every run. Nobody in the literature we found measures the yield of an individual query, which is exactly why we measure it on the screen. The quality report is not a thought-leadership position. It is the run log, itemised.

What we do not know

Two honest gaps. We do not benchmark your yield against other workspaces, so "typical kept rate" is not a number we publish; the right yield for your searches depends on your market's noisiness, and a universal average would be decoration. And we do not pretend the automatic disable rule is optimal for every listener. Three runs is a fixed cutoff, not a setting; a listener watching a slow market may want a longer leash, and for now the way to give it one is to save the plan again, which brings the pruned queries back.

If you want to see per-query yield on your own searches rather than ours, the free report runs a capped, real scan against your market, and a demo walks through the quality report. Either way, ask to see the rejections. The tool that shows you what it threw away is the one you can trust with what it kept.

References

• "Google Alerts for Company Tracking: Effectiveness & Alternatives", Contify, updated around September 2026. Vendor research. Source for the 10% relevance, 40% missed, and 92% survey figures.
• "X API Goes Pay-Per-Use: What Indie Hackers Should Build Now", Superframeworks, October 2026.
• "Reddit API Pricing 2026: Official Rates, Free Tier Limits & the $12K Minimum", Prowlo, October 2026.
• "Media Monitoring Solution Cost (2026 Guide)", Brand24, updated 28 April 2026.
• "Alert Fatigue: A Guide to Understanding and Reducing It", Workato, October 2026.
• "Buying signals are everywhere, most teams ignore them", Lusha, August 2026.
• "Automate data-driven alerts for B2B sales and success", Clearbit.
• "Social Listening ROI: The Metrics That Actually Prove Business Impact", Syncly, March 2026.
• "How to Measure ROI of AI Social Listening Tools (2026)", Digiflute, 2026.

Frequently asked questions

Next

• Start the free trial — Creates a workspace. A card is taken up front and nothing is charged during the trial.
• Pricing — What it costs after the trial.
• All posts — The comparisons and the guides, grouped.

See also

• What Open Pulse is and what it watches
• llms.txt — the index of this site, written for agents
• api.md — the MCP server and the REST API, with scopes
• auth.md — how to authenticate
• openapi.json — the REST API as OpenAPI 3.1
• sitemap.xml — every indexable page