Gap Radar
FSL-1.1-MIT, source available
Reads public complaints, reviews and forum posts, then writes new business ideas every week, each with linked quotes, demand signals and a score.
Run it
npm test node cli.mjs run
Released 18 Nov 2026 ยท Run 19 Nov 2026, re-checked with Gap Radar 3.1.0 on 20 Nov 2026: 0 ideas published, 6 leads, 2 rejected.
Top leadAccounting software for small firmsPart-time product
Feed Drop Sentinel
What it does
An automatic market researcher that runs once a week.
The full description
It picks the markets worth a look from a map of 244 (or takes the ones you list and the topics members vote for), reads real complaints, reviews and forum posts (Hacker News, Stack Exchange, GitHub issues, App Store and Google Play reviews, Lemmy, forums you list; G2, Capterra, Trustpilot, Reddit, Quora and Product Hunt through data vendors on your own keys) plus demand signals, and writes new business ideas in three tiers: bigger venture, part-time product or SaaS, side hustle.
It buys a small complaints-first sample per market first and spends the rest only on the 3 most promising problems, keeping every review it bought so none is paid for twice.
Every idea carries 3+ linked quotes with dates, labelled demand signals (count, proxy or real volume), the competition and its gap, a 1-day / 1-week / 1-month MVP, a first-10-customers plan, risks and a score on a 76-criterion rubric (10 radar families, gates and caps shown).
An AI analyst panel (Claude, Kimi, OpenAI, Ollama or any CLI, on your key or a plan you already pay) judges the criteria data cannot measure, with a hard cost cap; without a model it falls back to heuristic mode, judges fewer criteria and every card says so.
From the run's report: Gap Radar 3.1.0, scoring 2.1.0
No idea cleared the bar this run
Gap Radar 3.1.0 published 0 ideas from this run: 6 problems are leads and 2 were rejected. The bar went up, the scores were not tuned. Traced below, step by step, is the top-ranked lead, Feed Drop Sentinel (Accounting software for small firms, Part-time product), from the run's report.
- Step 1 of 6
The problem
Written by the AI model from the quotes
QuickBooks reviewers report occasional sync issues with bank feeds and third-party integrations that add friction. Xero reviewers say feeds sometimes disconnect or update slowly, and one reports a bank feed that dropped silently. GitHub write-ups describe SMB owners reconciling manually in spreadsheets that do not proactively flag errors, and finance teams re-entering platform fees and payments into accounting software by hand.
Card text by the AI model. Every sentence was split into claims and checked in code against the evidence ledger (quotes word for word, numbers, the product each quote is about), then by a second model where one is configured. Failed claims were removed and are listed on the card.
- Step 2 of 6
The evidence
Counted from the collected posts
4 first-hand voices back this problem. 6 supporting quotes: 4 first-hand, 2 second-hand, from 2 sources (Capterra review, GitHub issue), dated 3 Mar 2026 to 19 Nov 2026. Of 923 posts read in this niche, 85 were complaints.
The evidence check: 6 quotes read against this card's own problem by Claude Opus 5.5
- supporting4 first-hand, counted; 2 second-hand, shown only
- 6
- neutralleft out
- 0
- contradictingcounter-evidence, not counted
- 0
Pain: 0 of 4 first-hand voices intense, weighted by review rating and words; ratings 1 at 3 stars, 2 at 4 stars, 1 at 5 stars; mild words: sometimes, occasional, doesn't happen all the time, minor.
The 6 quotes, word for word, and the counter-evidence
The 6 supporting quotes on the card, word for word
Occasional sync issues with bank feeds or third-party integrations add friction.
Capterra review, 4 stars, 19 Oct 2026 (opens the original post in a new tab)
Current workaround: SMB owners use spreadsheets or basic accounting software to reconcile, but these require manual input and don't proactively flag likely errors.
GitHub issue, 3 Mar 2026, second-hand, not counted as a voice (opens the original post in a new tab)
For my sol prop it missed speratic transactiins and never lost connection to my account.
Capterra review, 3 stars, 25 Apr 2026 (opens the original post in a new tab)
Context: Finance teams re-enter platform fees and recruiter payments into their accounting software by hand.
Sometimes the feeds disconnect or take a long time to update, which slows down reconciliation. It doesn't happen all the time, but when it does, it's annoying because it breaks the smooth workflow Xero is known for.
Capterra review, 4 stars, 16 Sep 2026 (opens the original post in a new tab)
I did have a bank feed drop silently last March.
Capterra review, 5 stars, 6 Oct 2026 (opens the original post in a new tab)
Counter-evidence, shown and not counted
None: no checked quote contradicts this card (0 of 6).
- Step 3 of 6
Demand signals
Counts from the posts; indirect signs marked proxy
Step 3, Demand signals: 4 first-hand complaints; no search volume measured
Demand signals for this card, each labelled data or proxy Signal Kind 4 first-hand complaints, 2 second-hand write-ups 2 independent sources 28 points, likes and replies 0 posts say they would pay 1 in the last 7 days (too few for a rate) Search volume: not measured, so the score is capped at 80K6 none Data means a real count: posts we collected, or items an API reports (for example Hacker News mentions in 30 days). Proxy means an indirect sign of interest: search suggestions or Wikipedia page views. Useful for direction, never a search volume.
Named in the complaints
- Xero 3 mentions, 0 low-star
- QuickBooks 2 mentions, 0 low-star
- Step 4 of 6
The AI analyst
Judged by the AI model
Step 4, The AI analyst: Verdict: Watch. 2 of 3 judges said test first; the evidence check held it at Watch (4 first-hand voices, an idea needs 5).
The model judged 23 of 76 criteria in 3 independent passes; the median is kept and the spread between passes is shown. Every judgment cites evidence ids from the pack it was given.
Final verdict Watch
3 judges voted: 2 test first, 1 watch. The vote said test first; then the evidence check found 4 distinct voices (an idea needs 5), so the card is a lead, held at watch.
One judge's note, from the majority that said test first Worth a cheap test.
3 of its judgments, by weight
CriterionScore of 5SpreadConfidenceThe model's reason and evidenceSwitching costs and workflow embeddingO3300.35Books, invoices and bank feeds live inside the incumbent tool, and one voice cannot even export invoices they generated, so stored data and integrations raise the cost of leaving.Evidence capterra:0d0a85f5c350 (opens the post in a new tab)capterra:06e110697bc6 (opens the post in a new tab)capterra:872d4fd64487 (opens the post in a new tab)Current spend evidenceW2100.50Users clearly run paid subscriptions to QuickBooks, Xero and Wave, but no voice states they pay or names an amount.Evidence capterra:06e110697bc6 (opens the post in a new tab)capterra:872d4fd64487 (opens the post in a new tab)capterra:17c85bdb662f (opens the post in a new tab)Solution crowding (inverted)C3100.25Keyword-level counts show over a hundred listings for accounting software. The exact feed-monitoring job was not counted.Evidence measured C3Red team, a second AI pass that argues against the card. Its verdict: test first
Unverified model notes. Printed as the model wrote them. Gap Radar 3.1.0 checks that each sentence about a reviewer matches a review it cites, not whether the argument holds.
- K8One reviewer who reported a silent drop still rated the product 5. Another QuickBooks Online Accountant user says the connection was never lost.Evidence capterra:06e110697bc6 (opens the post in a new tab)capterra:17c85bdb662f (opens the post in a new tab)capterra:a80e433bc086 (opens the post in a new tab)
- K5The two GitHub items are feature-request write-ups that describe what SMB owners and finance teams do, not first-hand users.Evidence github:72d0202eda64 (opens the post in a new tab)github:e2ddf1fdad48 (opens the post in a new tab)
- Step 5 of 6
Validation
The model checked against the data
Step 5, Validation: Data and model agree on 2 of 4 shared criteria; pain intensity and incumbent dissatisfaction differ by 2 or moreP1, C1
4 criteria were read by both the data and the model. 2 agree within a point; 2 are flagged because they differ by 2 or more, and the card keeps both readings.
Criteria read by both the data and the AI model Criterion Data Model Result Pain intensityP1 1 3 flagged, 2 apart Budget holderM5 4 4 agree Incumbent dissatisfactionC1 3 0 flagged, 3 apart Gap specificityC4 2 3 agree Gates that fired
- No proven pain: Problem 1.93 of 5. Caps the score at 45.G1
- Single source family: 1 voice family with 3+ voices (1 in all). Caps the score at 60.G2
- Independence and span: 1 independent family, span 262 days, 1 in the last 30 days. Caps the score at 50.K2
- No real volume: no measured search volume. Caps the score at 80.K6
- Lead only: after the evidence check, 4 distinct voices (an idea needs 5). Caps the score at 40.LEAD
The lowest cap that fired is 40 (lead only). The unrounded score of 33.9 sits under it, so no cap changed the number.
An AI analyst panel scored the judgment criteria, gave a verdict, a tier and a niche fit per problem, citing evidence ids. Verdicts need 2 agreeing judges; criteria 2 points apart across judges are marked "models disagree" at confidence 0.25.
The evidence check, card by card
Before the check, 2 of 6 cards in this run were ideas. After it, none is. The bar went up, the scores were not tuned.
CardBefore the checkAfterQuotes checkedFeed Drop SentinelAccounting software for small firmsidea, 35, 6 voiceslead, 34, 4 voicesAfter the evidence check, 4 distinct voices (an idea needs 5).6 of 6 support, 2 of them second-handFlat-Price BooksAccounting software for small firmslead, 35, 9 voiceslead, 34, 7 voicesThe panel says watch: worth tracking, not building yet.7 of 9 support; 2 neutralSubscription cost cutterSmall landlordslead, 35, 5 voiceslead, 33, 3 voicesThe panel says watch: worth tracking, not building yet.4 of 5 support, 1 of them second-hand; 1 neutralFlat Rent BillingSmall landlordsidea, 34, 8 voiceslead, 33, 3 voicesAfter the evidence check, 3 distinct voices (an idea needs 5).3 of 8 support; 5 neutralOnboarding Setup ServiceAccounting software for small firmslead, 33, 16 voiceslead, 33, 13 voicesThe panel says watch: worth tracking, not building yet.13 of 16 support; 3 neutralCalm Accounting PlatformAccounting software for small firmslead, 33, 5 voiceslead, 33, 4 voicesThe panel says watch: worth tracking, not building yet.4 of 5 support; 1 contradictsAfter the check, voices are first-hand supporting voices only. Each reason is the report's own.
Counter-evidence, shown and not counted
If you want accounting software that works, has great integrations, isn't bloated with a lot of things you don't need, and doesn't try to sell you something every time you open the software, try Xero.
Capterra review, 5 stars, 24 Aug 2026 (opens the original post in a new tab)
Against Calm Accounting Platform: Says Xero is not bloated and does not upsell.
- Step 6 of 6
The score
The data and the model, combined
34of 100
Band weakLow confidenceRank 1 of 6
- Formula
- Score = opportunity 41.5 x (0.7 + 0.3 x confidence 0.39) = 33.9. Lowest cap that fired: 40 (lead only), not reached
- Status
- Lead, not published. After the evidence check, 4 distinct voices (an idea needs 5).
Why it ranks 1 of 6, its confidence, and what would change it
- Why it ranks here
- Rank 1 of 6 cards in this run, 0 published and 6 leads. Ranked 1 of 4 in this niche (verdict watch, score 34, part-time weights). Above "Flat-Price Books" because it scores higher on incumbent neglect (+0.2 points), a simpler MVP (+0.2 points), incumbent dissatisfaction (+0.2 points). Tied at 34 with "Flat-Price Books"; the tie is broken by the higher confidence (0.39 against 0.37). Order: an idea before a lead, then the stronger verdict, the higher score, the higher confidence, more first-hand supporting voices, fewer counter-evidence quotes, then the card id. All 5 caps that fired sit above 33.9, the lowest at 40 (lead only), so none changed the number.
- Confidence
- Low: lead, after the evidence check, 4 distinct voices (an idea needs 5); confidence factor 0.39: evidence quality 2.84 of 5, 56% of the weighted criteria assessed; forced low by the single source family gate (1 voice family with 3+ voices (1 in all))
- What evidence would change this
- Complaints from a third independent kind of source (today 2): one more source family.
- One person saying what they pay today or would pay for a fix, in their own words.
- Measured monthly searches for the problem in plain words ("accounting software syncing"), from a keyword tool.
- Search intent: connect Keyword Planner, Bing Webmaster or DataForSEO
- Platform or marketplace channel: count marketplace listings for the job
The 10 families behind each score
Every card is scored 0 to 5 on ten rubric families. Nine add up to the opportunity; the tenth, evidence quality, sets the confidence. A family with nothing assessed stays at the 2.0 prior.
All 6 cards in this run are band weak (under 40 of 100), all with Low confidence, and none was published. They score 33 to 34, 1 point apart: too close for their families to show what separates a card. So the chart draws a contrast the data does hold, from the backtest: Plausible Analytics, which reached revenue, against Quibi, which failed, both scored only from posts dated before they launched.
What separated a winner from a failure
Plausible Analytics (35 of 100, reached revenue) minus Quibi (32 of 100, failed), score of 5 per family
- Plausible Analytics ahead
- Quibi ahead
| Family | Plausible Analytics, Quibi | Difference |
|---|---|---|
| Problem | 2.45 vs 2.13 | +0.32 |
| Market | 2.26 vs 2.05 | +0.21 |
| Timing | both at prior | 0.00 |
| Competition and gap | 2.40 vs 2.11 | +0.29 |
| Money and willingness to pay | 2.00 vs 2.00 | 0.00 |
| Distribution | 2.09 vs 2.11 | -0.02 |
| Feasibility and AI fit | 2.00 vs 2.06 | -0.06 |
| Moat | 2.07 vs 2.00 | +0.07 |
| Founder fit | both at prior | 0.00 |
| Evidence quality | 2.87 vs 2.58 | +0.29 |
Calibrated on 23 real launches
- historical launches, 12 that reached revenue and 11 that failed
- 23
- ranking AUC, Gap Radar 3.1.0 (0.826 at calibration, 0.674 before the fix)
- 0.848
- best-ranked cases were winners
- 7 of 7
AUC is the chance that a real winner ranks above a real loser: 0.5 is a coin flip, 1.0 a perfect ranking.
How the backtest was run, and its limits
How the backtest was run
Before these scores were published, the scoring (rubric 2.1.0) was backtested on 23 problems from before launch, among them Plausible, Bitwarden, Linear and Figma, and Quibi, Juicero and Google Wave. Each was scored from Hacker News, Stack Exchange and GitHub posts dated in the 2 years before it launched, collected once on 19 Nov 2026; Gap Radar 3.0.0 and then 3.1.0 re-scored the same posts on 20 Nov 2026.
Every backtest case still scored weak and Low. The bands were not lowered to make winners pass. A test keeps the ranking honest: it fails if the AUC drops under 0.776, the 0.826 reached at calibration minus a 0.05 margin. Gap Radar 3.1.0 scores 0.848, counting first-hand voices only.
What this backtest cannot show
- Small sample 23 cases can show direction, not fine weights.
- Picked by hand the cases are well-known products, not a random draw of all ideas.
- Queries written knowing the outcome worded as each founder would have searched before building, for winners and losers alike, by someone who knew how each one ended. Other wording would collect other posts.
- Data only heuristic mode assessed 9 to 37% of the weighted criteria. Feasibility, moat, founder fit and most of market and timing sat at the 2.0 prior.
What it costs to run
Two ways to run it
Free, run by hand
How Source Included runs it.
- What it costs
- EUR 0 a month on top of the AI plan you already have.
- AI
- Your Claude Code plan, plus Kimi Code as a second judge if you have it. The calls count toward those plans' limits.
- Data
- The free sources, plus free tiers: Apify Free (USD 5 of use a month, about one run of G2, Capterra, Trustpilot, Reddit and Quora reviews), the Outscraper and Bright Data free allowances, and free GitHub, Stack Exchange and YouTube keys. Search demand comes from Google, Bing and YouTube suggestions, not measured volumes.
- When it runs
- When you start it, on your own signed-in computer.
- Trade-off
- Fewer bought reviews a run, so fewer problems reach 5 first-hand voices: more leads, fewer ideas. Reddit, Quora and Product Hunt only while the Apify free credit lasts.
Maximum
- What it costs
- EUR 20-21 a month on your AI plans (Apify Starter and DataForSEO), or EUR 46-55 a month fully in the cloud with the laptop closed (Kimi and Claude APIs, Apify, DataForSEO and a small server).
- AI
- Your Claude Code and Kimi Code plans, or the Kimi and Claude APIs on a server.
- Data
- Everything in the free setup, plus a paid sample of 20 reviews per review site per market and deep buying on the 3 most promising problems: about 1,800 bought reviews a run. Measured monthly search volumes and App Store, Google Play and Trustpilot reviews through DataForSEO.
- When it runs
- Every Monday at 06:00 on your computer, or around the clock on a server.
- Trade-off
- It costs money, and the cloud route needs API keys and a server.
Both run the same code, the same 76-criterion scoring, the same strict gates and the same audit; only the amount of data differs, and every report says which sources it used.
A weekly run is never free at full depth: every run pays for model calls and for the review and search data it buys. Totals a month, models plus vendors, in EUR, from the low to the high end of the likely range.
| Route | Lean4 niches | Funnel5 niches | Standard8 niches | Max12 niches |
|---|---|---|---|---|
| Cloud mixedruns 24/7 on a server: Kimi sorts and takes 2 analyst passes, Sonnet 5 takes the third and checks, Opus 5.5 writes | 36-43 | 41-49 | 62-78 | 165-221 |
| Cloud all ClaudeHaiku 4.5 sorts, Sonnet 5 judges, Opus 5.5 writes the cards, through the Batch API with prompt caching | 43-51 | 51-63 | 79-101 | 171-227 |
| All Sonnet 5every step on Sonnet 5 through the Batch API | 46-56 | 53-66 | 88-112 | 183-244 |
| All Opus 5.5every step on Opus 5.5, one call at a time, no batching | 127-166 | 154-202 | 291-387 | 430-576 |
| Claude Code plan and Kimi Codevendors only: the model calls count toward plans you already pay, on your own signed-in computer | 19-21 | 20-21 | 20-22 | 23-32 |
On a phone the table shows the configured Funnel profile; the other three are on a wider screen and in docs/COSTS.md.
The four run profiles, and the Funnel line by line
Funnel is the configured weekly run: 5 niches, a 20-item paid sample per source, complaints first, then deep data on the 3 most promising problems, with every review bought once into a shared library. Lean judges up to 10 problems in each of 4 niches. Standard buys deep data for every problem in 8 niches. Max runs 12 niches with Opus 5.5 as the judge. The cloud routes add a small server, EUR 4-6 a month.
Funnel on the Cloud mixed route, line by line, and what keeps it low
- AI modelsVariable
- EUR 21-29
- Relevance, review sorting, job grouping, 3 analyst passes per problem on 2 models, red team, cards, evidence check, verifier, topic scout
- Apify StarterFixed
- EUR 17.50
- USD 19 a month with USD 19 of use included: G2, Capterra, Reddit, Trustpilot, Quora, Product Hunt
- DataForSEOVariable
- EUR 2-4
- Search volume and the App Store, Google Play and Trustpilot review routes, paid from a USD 50 prepaid balance that does not expire
- Cloud serverFixed
- EUR 4-6
- A small Linux server, so the weekly run goes on with the laptop closed
- Free sourcesFree
- EUR 0
- YouTube Data API, GitHub, Stack Exchange, Hacker News, Lemmy, the App Store feed
What keeps it low
- Sample before buying: 20 items per source per niche, complaints first, then deep data only on the 3 problems with the most promise, and it stops early once a problem has enough evidence.
- Bought once: every review goes into a library shared across niches, so later weeks fetch only what is new (about 1,800 paid items a week once it fills, from about 3,780 in week 1).
- Rules before models: documents that name no niche keyword, competitor or audience word leave before any model reads them.
- The cheapest model that can do the step, the Batch API at half price, and the analyst's fixed instructions read from the prompt cache.
- Hard caps: vendors at USD 5 a run and USD 24 a month, 25 problems a niche, 3,000 documents a niche.
The cloud routes need an Anthropic API key, a Kimi API key for the mixed route, and a small server. Until those are set up, the weekly run uses the Claude Code plan route on the owner's computer.
From the cost model in the Gap Radar repo (docs/COSTS.md), calibrated on the stored model calls of the 19 Nov 2026 example run. Ranges are p10 to p90 a month over 4.33 weekly runs, at 1 USD = 0.92 EUR. Estimates, recalibrated after the first weekly runs.