# Head-to-head: mathtec/google-trends (ours) vs apify/google-trends-scraper (Store leader), 2026-11-09

Bead: hustle-qtk. Purpose: find out whether our Store listing can honestly say "better", and if so where.

Leader confirmed from the public Store API on 2026-11-09 (search "google trends", sorted by 30-day users): **apify/google-trends-scraper, 3,272 users**. Second is data_xplorer/google-trends-fast-scraper with 895 users (`h2h-google-trends/actors/store-search-2026-11-09.json`). Ours had 0 users in 30 days.

## Outcome first

- **Data: equal, within Google's own sampling noise (n=18 keyword×geo pairs with data).** The interest-over-time series agreed point for point on 8/18 pairs. On the other 10 the median correlation was 0.997 (median mean absolute difference 1.04 on the 0–100 scale; max difference 10 on high-volume terms, 100 on the low-volume "bitcoin sv"). Our own Actor, run twice two minutes apart, also differed from itself on 15/18 pairs, with up to 9 points on a high-volume term (mortgage rates, US). Leaving out the still-partial current week, **the leader's series was identical to one of our two passes on 11/18 pairs**, and our two passes were identical to each other on 9/18. Google serves a few different samples. Both Actors pass through whichever sample they get, and neither is more accurate.
- **Low-volume terms show the noise at its largest.** On "bitcoin sv", our pass 1 and the leader differed by up to 100 points in GB and 50 in US (r 0.24 and 0.83). Our pass 2 was **identical** to the leader on both geos (53/53 points each).
- **Success: equal on terms with data.** 18/18 each. "hodlonaut" (very low volume) came back from both as no data in US. In GB ours returned no series twice, while the leader returned a series with one week at 100 and the rest 0. That is another sampling difference, not one we can call either way.
- **Price: ours is lower.** On this STARTER (BRONZE) account: ours **$0.0014 per keyword result**, with no start fee; leader **$0.00505** ($0.005 per result + $0.00005 start). So ours costs **0.28x as much** (n=18). On no-data keywords, ours charged **$0** (2/2) and the leader charged **$0.00505** (2/2).
- **Speed: ours faster.** Run seconds per single-keyword run: median **6.9 s ours vs 28.5 s leader** (n=20 each), p90 17.2 vs 48.4, max 34.7 vs 55.9. Our second pass had a median of 4.5 s.
- **Fields: mostly equal, with differences each way.** Both returned the same region rows (4 GB nations, 51 US states) and the same counts of top and rising related queries on the high-volume terms. The related-query lists were identical on 8/18 pairs, and the rest differed by sample. Related topics came back empty from both on 20/20 (ours says so in `notes`). **Leader only:** `interestByCity` (up to 200 rows) and, for US, `interestByMetro` (210 DMAs) in the same row. Ours returns one region resolution per run (`regionResolution`). **Ours only:** `summary` (average, peak, peak date, latest, latest-is-partial), `status`, `hasData` and plain-language `notes` ("Google returned no relatedTopics for …"). Ours lists only the regions Google returns with data, while the leader also lists zero-value regions (bitcoin sv US: 21 rows vs 51).
- **Failure / no-data handling: ours clearer.** On a no-data keyword, ours writes a row with `hasData: false` and notes, and does not charge. The leader writes a full row of zeros with no flag beyond per-point `hasData`, and charges for it.

**Claims the numbers support for the listing:**
- "Same Google Trends data as the category leader at 28% of the price" (n=18; differences are within Google's own sampling, shown by running ours twice).
- "Keywords with no Google data are not charged" (2/2; the leader charged 2/2).
- "Faster per keyword": median 6.9 s vs 28.5 s on single-keyword runs (n=20).
- "Summary per keyword (average, peak, latest)" (a field the leader lacks).

**Claims the numbers do NOT support:**
- "More accurate" or "more complete": the series are the same up to Google's sampling.
- "More data per row": the leader returns city and metro rows that ours returns only on request, at one resolution per run.

**data_xplorer/google-trends-fast-scraper (#2, 895 users): its current build ignores the keyword.** On 4/4 runs (bitcoin GB, bitcoin US, bitcoin sv GB, hodlonaut US), it returned Google's **US "Trending now" list for the past 24 hours (50 items)**, not the requested keyword's 12-month interest. Its log says "Configuration for Trending Searches … Region: US", even for `geo: GB`. It charged **$0.026 per run** ($0.025 start + $0.001 per item). This is one build (2.10.2) on one day. It is not a claim about the Actor's normal behaviour, but on 11-09 it did not return keyword data.

## Method

1. **Inputs.** 10 keywords × 2 geos (GB, US) = 20 pairs, past 12 months, weekly (53 points). There are 8 high- or medium-volume terms, plus 2 low-volume terms from the org's Craig lane ("bitcoin sv"; "hodlonaut", expected to be near zero). The list and reasons are in `inputs.json`.
2. **Runs.** All runs came from the Hustle org (`mathtechustle`, STARTER/BRONZE). This org does not own `mathtec/google-trends`, so it pays the public price. Each pair got one single-keyword run per Actor, started at the same moment (5 pairs, so 10 runs, at a time). Every run was async (`POST /v2/acts/<id>/runs`, then poll) with `maxTotalChargeUsd` ($0.01 ours, $0.02 leader, $0.04 data_xplorer) and a client guard that aborts above 5 dataset items. Default memory and build; no Actor settings or prices changed.
3. **Noise pass.** About 2 minutes after pass 1, ours ran all 20 pairs again (`runs/ours2`). Google returns slightly different samples to repeated requests, so this pass is the noise floor for comparing ours to the leader.
4. **Equivalent settings** (exact inputs are in each `runs/*/*.json`):

   | Actor | input |
   |---|---|
   | ours | `keywords: [kw]`, `geo`, `timeframe: "today 12-m"`, `interestOverTime: true`, `interestByRegion: true`, `relatedQueries: true`, `relatedTopics: true`; the rest default (`compareTogether: false`, `category: 0`, `property: ""`, `regionResolution: ""`, `language: en-US`, `proxyStrategy: datacenter-then-residential`) |
   | leader | `searchTerms: [kw]`, `geo`, `timeRange: ""` (its value for "Past 12 months"), `maxItems: 0`; the rest default. It returns every data block by default, so ours had all four of its blocks switched on to match. |
   | data_xplorer | `keyword`, `geo`, `timeframe: "today 12-m"`, `language: en-US` (its whole schema) |

5. **Reference.** Google Trends has no independent ground truth: every client gets a sample, and the values are scaled 0–100 per request. The reference is therefore the comparison between Actors **and** ours against itself. Series are compared by date: points identical, mean and max absolute difference, and Pearson r. The current week is partial (`isPartial`) and moves between requests, so the "identical to one of our passes" figure leaves it out.
6. **Metrics.** Success means a series with at least one non-zero week. $ is `usageTotalUsd` per run after the runs settled. All 104 run objects in this bead were re-read, and $ equals charged events × event price on every one. Seconds are `stats.runTimeSecs`.

## Results

| Actor | runs | with data | $ total | $ per input | $ per result | median s | p90 s | max s |
|---|---|---|---|---|---|---|---|---|
| ours (pass 1) | 20 | 18 | 0.0252 | 0.00126 | 0.0014 | 6.9 | 17.2 | 34.7 |
| ours (pass 2) | 20 | 18 | 0.0252 | 0.00126 | 0.0014 | 4.5 | 6.9 | 21.7 |
| leader | 20 | 19 (incl. hodlonaut GB, 1 non-zero week) | 0.101 | 0.00505 | 0.00532 | 28.5 | 48.4 | 55.9 |
| data_xplorer | 4 | 0 (wrong data: US trending list) | 0.104 | 0.026 | - | 19.3 | 30.1 | 60.4 |

Series agreement over 18 pairs (954 points):

| comparison | points identical | pairs fully identical | median of per-pair mean abs diff | max abs diff | median r |
|---|---|---|---|---|---|
| ours vs leader | 643 / 954 | 8 | 0.67 | 100 (bitcoin sv GB) | 0.998 |
| ours pass 2 vs leader | 644 / 954 | 5 | 0.02 | 10 | 1.000 |
| ours vs ours pass 2 (noise) | 690 / 954 | 3 | 0.03 | 100 (bitcoin sv GB) | 1.000 |

Excluding the partial current week: the leader was identical to ours pass 1 or pass 2 on **11/18** pairs, and ours pass 1 = pass 2 on **9/18**.

## Per pair

Related = top + rising related queries returned. "weeks >0" counts non-zero weeks out of 53.

| keyword | geo | volume | weeks >0 ours / ours2 / leader | ours vs leader: identical pts, mean abs diff, max, r | ours vs ours2: identical pts, mean abs diff, max | regions ours / leader | related top+rising ours / leader | $ ours / leader | run s ours / leader |
|---|---|---|---|---|---|---|---|---|---|
| bitcoin | GB | high | 53 / 53 / 53 | 19/53, 0.74, 3, 0.998 | 5/53, 1.87, 4 | 4 / 4 | 25+25 / 25+25 | 0.0014 / 0.00505 | 6.0 / 31.8 |
| bitcoin | US | high | 53 / 53 / 53 | 23/53, 0.68, 2, 0.998 | 22/53, 0.81, 2 | 51 / 51 | 25+25 / 25+25 | 0.0014 / 0.00505 | 17.2 / 27.0 |
| chatgpt | GB | high | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 52/53, 0.02, 1 | 4 / 4 | 21+12 / 21+12 | 0.0014 / 0.00505 | 5.0 / 25.6 |
| chatgpt | US | high | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 53/53, 0, 0 | 51 / 51 | 24+16 / 24+16 | 0.0014 / 0.00505 | 3.5 / 37.2 |
| mortgage rates | GB | high | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 52/53, 0.02, 1 | 4 / 4 | 25+25 / 25+25 | 0.0014 / 0.00505 | 7.1 / 25.7 |
| mortgage rates | US | high | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 9/53, 1.92, 9 | 51 / 51 | 25+25 / 25+25 | 0.0014 / 0.00505 | 6.9 / 27.0 |
| electric car | GB | medium | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 53/53, 0, 0 | 4 / 4 | 25+25 / 25+25 | 0.0014 / 0.00505 | 34.7 / 48.4 |
| electric car | US | medium | 53 / 53 / 53 | 8/53, 1.81, 6, 0.998 | 52/53, 0.02, 1 | 51 / 51 | 25+25 / 25+23 | 0.0014 / 0.00505 | 7.4 / 55.9 |
| premier league | GB | high in GB | 53 / 53 / 53 | 18/53, 1.06, 4, 0.998 | 51/53, 0.04, 1 | 4 / 4 | 25+15 / 25+14 | 0.0014 / 0.00505 | 6.2 / 32.5 |
| premier league | US | lower in US | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 52/53, 0.02, 1 | 51 / 51 | 25+23 / 25+23 | 0.0014 / 0.00505 | 16.6 / 28.6 |
| taylor swift | GB | high | 53 / 53 / 53 | 20/53, 1.02, 10, 0.994 | 53/53, 0, 0 | 4 / 4 | 25+25 / 25+25 | 0.0014 / 0.00505 | 6.7 / 36.9 |
| taylor swift | US | high | 53 / 53 / 53 | 25/53, 0.66, 3, 0.999 | 31/53, 0.42, 1 | 51 / 51 | 25+25 / 25+25 | 0.0014 / 0.00505 | 8.0 / 32.2 |
| solar panels | GB | medium | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 52/53, 0.02, 1 | 4 / 4 | 25+25 / 25+25 | 0.0014 / 0.00505 | 5.4 / 24.8 |
| solar panels | US | medium | 53 / 53 / 53 | 24/53, 0.87, 6, 0.995 | 24/53, 0.87, 6 (pass 2 = leader) | 51 / 51 | 25+25 / 25+25 | 0.0014 / 0.00505 | 9.7 / 28.1 |
| inflation | GB | medium | 53 / 53 / 53 | 53/53, 0, 0, 1.0 | 52/53, 0.02, 1 | 4 / 4 | 25+21 / 25+13 | 0.0014 / 0.00505 | 6.0 / 27.1 |
| inflation | US | medium | 53 / 53 / 53 | 17/53, 1.23, 5, 0.995 | 12/53, 1.3, 6 | 51 / 51 | 25+25 / 25+25 | 0.0014 / 0.00505 | 25.0 / 28.4 |
| bitcoin sv | GB | low | 5 / 5 / 5 | 48/53, 6.3, 100, 0.244 | 48/53, 6.3, 100 (pass 2 = leader) | 1 / 4 (non-zero 1 / 2) | 1+1 / 1+1 | 0.0014 / 0.00505 | 6.8 / 25.6 |
| bitcoin sv | US | low | 34 / 36 / 36 | 17/53, 9.3, 50, 0.831 | 17/53, 9.3, 50 (pass 2 = leader) | 21 / 51 (non-zero 19 / 22) | 5+2 / 5+2 | 0.0014 / 0.00505 | 3.5 / 52.0 |
| hodlonaut | GB | low | 0 / 0 / 1 | ours: no series, `hasData: false`, not charged | - | 0 / 4 (all 0) | 0 / 0 | 0 / 0.00505 | 2.2 / 27.7 |
| hodlonaut | US | low | 0 / 0 / 0 | ours: no series, `hasData: false`, not charged | - | 1 / 51 (all 0) | 0 / 0 | 0 / 0.00505 | 8.6 / 29.6 |

## Spend

- **This test (attributed per run): $0.2554.** Ours $0.0252 + $0.0252 (second pass); leader $0.101; data_xplorer $0.104. The three h2h tests in this batch together came to $0.8734, against a hard cap of $1.50 shared across all three.
- **`/v2/users/me/limits`:** `monthlyUsageUsd` was 7.5482 before the batch (`limits-before.json`) and 8.4247 after (`limits-after.json`), a delta of **$0.8765**. The 104 run ids of the three tests sum to $0.8734. No other org runs started between the two snapshots (the test runs themselves ran 09:47–09:57Z), so the remaining $0.003 is platform usage that cannot be attributed to a run (storage or transfer).

## Limits of this test

- One noise pass, for ours only. A second leader pass would be needed to measure the leader's own run-to-run variation; the evidence that it has the same variation is indirect (ours pass 2 matching the leader).
- Single-keyword runs only. Neither Actor's multi-keyword batching or `compareTogether` was measured. The leader's 30 s median is mostly per-run start-up, so batches would narrow the speed gap.
- One timeframe (12 months) and two geos.
- data_xplorer was sampled on 4 pairs because of its $0.025 start fee per keyword run.

## Files

All under `/home/a/Projects/swarm/agents/hustle/reports/market/h2h-google-trends/`:
- `inputs.json`: the 10 keywords, 2 geos and reasons.
- `actors/*.json`: each Actor's metadata, pricing and latest-build input schema at test time; `actors/store-search-2026-11-09.json`: the Store search used to confirm the leaders for all three tests.
- `runs/{ours,ours2,leader,dx}/<keyword>_<geo>.json`: input, cap, settled run object (`run`, plus `run_at_capture`), dataset items, log and client timing.
- `per-pair.json`, `summary.json`: the computed metrics.
- `limits-before.json`, `limits-after.json`: account usage snapshots for the whole three-test batch.
- `runs-console-pass2-dx.log`: the console output for the noise pass and data_xplorer.
- `scripts/`: `apify.py`, `store_leaders.py`, `getact.py`, `runlib.py` (async run + poll + guard + resettle), `runner.py`, `analyse.py`. The token is read from `.secrets/apify-org.env` at run time and is never written to disk.
