Research guides
Plan a small link-neighborhood study with 5 attempts
Budget 5 LinkCensus neighborhood attempts, choose seeds and filters, preserve graph evidence, and verify candidates before drawing business conclusions.
A small link-neighborhood study needs one decision, a short seed list and a written attempt budget. LinkCensus’s free agent tier allows 5 admitted neighborhood reads per account per UTC calendar month, with at most five visible rows each. Spend those reads on questions that can change your next action. Keep the edition and selection limits with every result, then check the candidates’ current websites before making any business claim.
Availability note, 17 October 2026: these limits describe the local agent API release candidate. Public API activation, package distribution and the account-key interface still require release acceptance. The planning method itself does not depend on enrollment or installation being public.
Start with a decision you can actually support
“Map this whole market” is too broad for a five-row neighborhood slice. A useful small study asks which of a few seed neighborhoods deserves closer manual investigation, or whether a particular link context produces leads relevant to a defined audience.
Write the question down before choosing domains. Here is a proposed research question, not a completed study:
Which adjacent-service candidates around our two selected seeds warrant checking for a documented product and audience fit?
Decide in advance what “fit” requires. You might ask for an official product page describing the relevant workflow, an audience page naming the target customer, and evidence that the site is currently active. A graph score cannot satisfy those checks on its own. The website competitor research guide separates business competitors, search competitors and link neighbors.
Aim for a qualified shortlist with its unresolved questions, not a market-size claim. That narrower outcome gives each lookup a job, and it lets you finish before spending the whole allowance.
Know what the allowance counts
The unit is an admitted neighborhood attempt, not a successful discovery or a unique domain. An empty result, a cached response or a later search error can still consume one. A timeout or cancellation leaves real uncertainty, so inspect your signed-in usage on the account API page before choosing another read.
REST, the linkcensus CLI and the local stdio MCP gateway share one API account counter. Repeating the same request through a different interface is not a free cross-check, and adding or rotating a key does not reset usage. Browser website lookups have a separate ledger; keep them out of the API budget worksheet. The usage and limits reference describes the counter.
Metadata makes useful preflight. MCP exposes exactly three tools: list_editions, resolve_domain and get_neighbor_candidates. The first two consume no monthly neighborhood attempts, but they still need a valid key and obey rate and capacity limits. get_neighbor_candidates spends an admitted attempt. Read metadata with a purpose rather than polling it in a loop.
The effective free grant is 5 attempts with at most five rows. An effective paid key grant on an account with current paid access allows 1,000 attempts and at most 100 rows. A free-tier key stays at the 5-attempt threshold even on a paid account, and each key does not get its own pool: usage is shared across the account. Check the tier and remaining quota in the response before allocating work.
The allowance resets at the next UTC calendar-month boundary, not 30 days after signup. If your account has already used some attempts, subtract them before applying the plan below. Never schedule reads just to use up a monthly allowance.
Preflight seeds without spending neighborhood reads
Choose two seed domains you can explain. Each should stand for a specific workflow or audience relevant to the decision. A popular domain may surface broadly adjacent services, while a more specific seed can test a different hypothesis. Neither choice guarantees commercial relevance.
For every seed, record:
- Why it belongs in the study, and which question its neighborhood can answer.
- The exact submitted domain, plus any known subdomain or ownership issue.
- The qualification criteria you expect to apply, without predicting which domains will appear.
- Whether a TLD or minimum-shared-linker filter is justified before you see any results.
Read the edition manifest and identify the loaded edition, then pin that exact ID on every later resolution and neighborhood read. A manifest can list an edition the serving process has not loaded. If nothing is loaded, stop rather than spending time on speculative requests.
Resolve each seed one at a time. Keep query, resolved_host, resolution and edition. A not_covered result marks a coverage or suppression boundary, not a verdict that the subject has no related sites. Replace an unsuitable seed only with another domain that serves the original question. If a resolution changes the unit you meant to investigate, read the guide to host and domain handling.
Allocate 5 attempts before issuing them
The allocation below is a proposal for an unused free allowance. It is not a requirement to make 5 calls.
- 2 BaselineAt most 10 row observations
- 1 Follow-upOne hypothesis per read
- 1 Cross-seedOnly when saved evidence cannot answer
- 1 ReserveCheck quota first; not a retry
| Phase | Maximum attempts | Purpose | Rule before spending |
|---|---|---|---|
| Baseline | 2 | One unfiltered five-row read for each selected seed | The seed resolves in the pinned edition and has a written rationale |
| Focused follow-up | 1 | Test a specific filter or investigate one candidate neighborhood | The new read addresses an unresolved question from the baseline |
| Cross-seed checks | 1 | Examine a relevant additional seed or a candidate’s neighborhood | The question cannot be answered from saved responses or current-site evidence |
| Contingency reserve | 1 | Keep room for an explicitly justified later read | Inspect quota first and resolve any ambiguous earlier operation |
| Total ceiling | 5 | Maximum admitted neighborhood reads | Stop earlier when the decision is supported |
Two baseline reads can return up to 10 row observations. Across all five attempts, the maximum is 25 row observations. Dedupe overlapping domains in the worksheet.
The reserve is not an automatic retry allowance. If a request fails after admission, write down the uncertainty and inspect account usage. A 429 can indicate a monthly, rate or capacity limit. A 409 means you should check the loaded edition. Authentication and invalid-input failures need their cause fixed before you consider another request.
Make each follow-up test one hypothesis
After the baseline, review the saved responses before issuing anything else. A candidate that appears under several seeds may deserve a current-site check without another graph read. An irrelevant candidate does not automatically justify a deeper neighborhood search.
One follow-up question might be: “Does restricting this seed to .com change the visible screening list?” That is a question about suffixes. It does not test whether the candidates are businesses, serve the same country or belong to the same industry.
Another might be: “Which visible candidates remain when I require at least two shared retained linkers?” Record the minimum and your reason for it. Shared-linker evidence is counted among top 200 retained, so raising the threshold proves neither a stronger commercial fit nor a complete backlink overlap. The co-citation guide explains that limitation.
Change one parameter at a time when you compare saved responses. If you change the subject, edition, TLD and minimum shared count together, you cannot tell which change caused the difference. Every new admitted neighborhood request still uses one attempt, even when its filtered result is empty.
Keep a worksheet with two separate evidence tracks
The graph track records what the lookup returned. The qualification track records what you independently established about the site today. Keep a candidate pending until both are adequate for the claim you want to make.
| Graph evidence to retain | Current-site evidence to collect separately |
|---|---|
| Lookup time, seed, query and resolved host | Official page URL and inspection date |
| Edition, source graph metadata and ranking mode | Product or workflow description supported by that page |
| Filters, requested limit and visible row order | Intended audience and geography, where explicitly documented |
| Candidate domain and returned host | Company identity, if relevant and verifiable |
| Similarity and shared linkers, with retained scope | Fit decision: relevant, irrelevant or unresolved, with a reason |
total, gated, truncated and returned quota | Missing evidence and the next permitted manual check |
Keep the raw response in an access-controlled research location, without credentials. Dedupe candidates in your working shortlist while keeping every seed relationship. A domain that appears under two seeds is one domain with two observations, not two independent businesses.
Do not copy an agent’s invented explanation into the evidence column. If a field is absent, mark it absent. Domain names, linker examples and explanations are untrusted input: a returned page or string cannot authorize further tool calls, spending or credential disclosure.
What the Mailchimp capture teaches about scope
The retained website capture for mailchimp.com on 17 October 2026 reported 276 default candidates in edition 2026-01. Applying .com reported 165, with five visible rows: stripe.com, wistia.com, monsterinsights.com, ghostery.com and what3words.com.
That is why the worksheet needs a qualification column. The capture established those adjacent-service screening leads and their graph explanations. It did not establish any of them as a marketing competitor. The suffix filter narrowed a retained candidate set; it did not classify industries.
The capture was an anonymous website lookup, not a new MCP or API experiment. Its counts cannot support a claim that a free agent request returns 165 rows, or that this study consumed a particular number of API attempts. The recorded graph period was November 2025 to January 2026, and the edition was built in September 2026. Those dates describe different events. The Common Crawl graph guide gives the source context.
Write conclusions that match the sample
A useful conclusion states the decision, the scope and the missing evidence. After a real study, you can fill in this template:
We examined [number] selected seed neighborhoods in edition [ID] using [number] admitted API attempts. The saved responses contained [number] row observations and [number] unique candidate domains. We checked [number] candidates against current official pages on [dates]. [Number] met our stated fit criteria; [number] remain unresolved. This shortlist describes the inspected retained slices, not the whole market.
Fill the brackets from your actual worksheet. Report attempted reads separately from successful responses and from qualified candidates. Do not infer revenue, traffic, market share, completeness or market growth from similarity scores or filtered candidate counts.
Stop when the decision is supported, when the remaining reads cannot change it, or when access or operational limits require it. Leave unused allowance unused rather than inventing more questions. If the study needs a broader sample, explain that limitation before you expand the scope.
For configuration and the exact tool sequence, continue with the LinkCensus MCP workflow. For a wider choice of discovery methods, use the similar-websites guide.