SEO Data Bench/Guides/serp-keyword-clustering-by-route

Does SERP keyword clustering depend on the SERP API? 50 keywords, six runs, measured

Field guidecohort v1US · enrun 2026-09-28includes a Treg route

The rule people give Claude Code with a SERP API: group keywords whose top 10 results share 5 or more URLs. Applied to the same 50 keywords through nine SERP routes across six runs between 15 September and 15 October 2026, how many keyword pairs each route groups, how far the groups move between runs half an hour apart, and what 1,000 keywords cost per route.

Written for people asking: How do I cluster keywords by SERP overlap with the DataForSEO API in Claude Code?Do SERP-based keyword clusters change depending on which SERP API you use?How much does SERP-based keyword clustering cost per 1,000 keywords?

The rule comes from a post on X that gives Claude Code the DataForSEO MCP and a Search Console query export: “pull the top 10 for every query (SERP API, standard queue), group keywords wherever 5+ of the 10 URLs match, that’s ONE intent” (@ecomtryggvi, 13 October 2026). The rule is sound and widely used. What it assumes is that the top 10 for a keyword is one list. On 28 September this site measured that it no longer was: the same query came back as two or three different result lists depending on the route and the call. This page measures what that does to the clusters.

What was run

Cohort v1 is 50 US-English buyer queries for data APIs, frozen on 15 September 2026 and listed on the dataset page. Their top-10 organic results were fetched on six runs: 15 September, twice on 28 September (about 30 minutes apart, the second time through four more routes), 5 October, 12 October and 15 October 2026. That is 1,700 stored responses. Fifteen October was a fresh run for this page: 250 calls through the five original routes, all of which returned data, for $1.14 off the Treg balance.

The scorer applies the rule above with no provider call. It normalises each URL, links two keywords when their top 10 share at least 5 URLs, and treats each connected set as one group. It then compares the links between routes on the same run and between runs on the same route. The thresholds of 3 and 4 shared URLs are in the tables as well; they move the counts up but do not change the pattern.

How many pairs each route grouped

Keyword pairs linked at 5 or more shared URLs out of 10, per route and run. Serpstat answers from a stored index and holds only 20 of the 50 keywords.

The first run on 28 September also went through four more routes: ScrapeCreators linked 15 pairs, LimaData 14, Serper 14 and AnyAPI 1. Across all eight live routes that run, 23 pairs were linked by at least one route, 7 by only one, and none by all eight.

Treg’s routed endpoint was served by DataForSEO on 15 September and by ScrapeCreators on every run since. Its column after 15 September is ScrapeCreators’ result list: on 28 September its links matched direct ScrapeCreators calls at 0.93, the share of links in common out of all links either found.

The same route, half an hour later

Four keyword pairs that any clusterer should group, with the number of top-10 URLs each pair shared. Five or more means grouped.

The other two pairs, “clay alternatives” with “clay alternatives for developers” and “meta ad library api” with “ad library api”, follow the same pattern; all four pairs for every route and run are in example-pairs.csv.

A singular and its plural sharing no result is not a change in intent. The likelier reading, given the 28 September depth check (one DataForSEO query returned two disjoint lists minutes apart), is that the two keywords were served from different lists. A clusterer cannot tell that apart from two intents, so it splits the group.

Measured as links in common with the same route’s previous run, at 5 shared URLs:

• DataForSEO: between 0 and 0.25 on every pair of runs since 28 September, including 0.20 between the two runs 30 minutes apart.
• Cloro and SerpApi: no link in common between any two runs from 28 September to 12 October. On 15 October both moved back towards 15 September: Cloro’s links matched its 15 September links at 0.75, SerpApi’s at 0.31. Their result lists still differed from each other’s (URL agreement 0.65) and from DataForSEO’s (0.31 and 0.24), so this is not the split ending.
• Treg’s routed endpoint (ScrapeCreators): between 0.67 and 0.93 on every pair of runs since 28 September.
• Serpstat: 1.00 on every pair, because a stored index does not re-fetch. It also holds fewer than half the keywords.

What 1,000 keywords cost per route

One top-10 fetch per keyword, at the rates measured on 28 September 2026 in the nine-route run:

The post names DataForSEO’s standard queue rather than the live endpoint used here. Google organic task-post is not listed in the Treg catalog, so its price and its lists were not measured; whether queued results split the same way is open. Checking a clustering run by fetching twice doubles every figure above.

Who this fits

• You cluster keywords from a SERP API in Claude Code or a script. Put one obvious plural pair from your own list (or “backlinks api” / “backlink api”) into the batch. If the run does not group it, the route served split lists that day, and the groups are not yet usable.
• You need groups you can repeat. Fetch the whole list twice, cluster both, and keep only the groups that appear in both runs. On DataForSEO in October that kept almost nothing; on the ScrapeCreators-served route it kept most groups.
• You use a routed endpoint. Record _treg.served_by with every fetch. The stable column above belongs to ScrapeCreators, not to routing; on another account, or another week, the routed call may be served by another provider. Pin one provider for a clustering job so every keyword in it comes from the same source.
• You want groups that never move. A stored index gives you that, for the keywords it holds: 20 of these 50, on data that does not refresh at request time.

What this run cannot tell you

Which grouping is right. The route with stable groups returns one of the lists Google is serving, and this run has no evidence on which list a person searching from the United States sees. It cannot say why the lists split or why Cloro and SerpApi moved back on 15 October. It is 50 keywords in one niche, in US English on desktop, with six runs, most of them a week apart. The 5-of-10 rule is the one from the post; tools that use hard clustering (every pair in a group must match) or a different threshold will produce different counts. The runs of 15 September to 12 October were made for other benchmarks on this site and reused here; only the 15 October run was made for this page.

Reproduce it

treg auth login
scripts/run-seo-cohort.py evidence/fixtures/seo-cohort-v1.json evidence/serp-runs/$(date +%F) --tasks serp
scripts/score-serp-clusters.py evidence/fixtures/seo-cohort-v1.json out/ 2026-09-15=public/evidence/seo-runs/2026-09-15 today=evidence/serp-runs/$(date +%F)

• Tables as CSV: pairs per route and run, groups at 5 shared URLs, route and run comparisons, example pairs, and the full summary. The tables hold keywords and counts, not result URLs.
• Raw responses for 15 and 28 September are on the evidence page; the scorer reads any directory of them.
• Scripts: cohort runner, clustering scorer, which imports the URL parsing of the SERP-check scorer.

Next step

• Run the scorer on two fetches of your own keyword list before acting on the groups. To pin a provider on Treg’s routed endpoint, the route headers are in the docs: treg.to/docs.
• Why the lists split, and what each route returned on 28 September: SERP APIs after Google’s scraper crackdown. Feeding a Search Console export to Claude Code in the first place: Search Console and keywords in Claude Code.

Relationship: the people who run this site also work on Treg, one of the routes compared on this page. Every Treg figure here is a call we paid for, with the raw response downloadable, and the same fixture is run against the alternatives where their terms allow.

Spotted an error or a changed price? Email corrections@seodatabench.com with the page URL and a source. Corrections are logged on the corrections page.