Click Decay MCP¶
The most expensive sentence in a traffic report is "our CTR dropped", said by someone who never split the change into the part that came from impressions. A site loses 1,000 clicks after a site migration and the room agrees on the cause: rankings fell, the snippets got worse, an algorithm update. The arithmetic that would settle it takes one line - clicks at the old CTR applied to the new impressions - and almost nobody runs it, because the person reporting is the person whose explanation depends on the answer. Click Decay is a free hosted server from a small Estonian research shop that runs that line for you and then labels what it just did: every number comes back as observed, calculated, modelled or not_established, and when Google's own Search Console notices cover the dates, the figures that depend on impressions come back null with a reason instead of a number.
Server type: remote Streamable HTTP, plain JSON responses, no session, keyless
Endpoint: https://clickdecay.com/mcp
Server: clickdecay 1.0.0 (serverInfo), protocol 2025-06-18, capabilities: tools only
Registry: com.clickdecay/mcp-server 1.0.0, active, remote declared as the same URL
Licence: MIT (github.com/meelis545/clickdecay-mcp, 0 stars at reading)
Vendor: QQQ OÜ, reg. 10975665, Tallinn - ai@qqq.ee
Tools: 4, every one annotated readOnlyHint true, idempotentHint true,
openWorldHint false
Limits: ~60 requests a minute per connection; takes numbers, dates and a
source name only - no queries, page URLs, site names or files
The evidence tier¶
This one is fully driven. A keyless initialize answered 200 with no authentication challenge of any kind, no session header, and capabilities carrying tools and nothing else - no resources, no prompts, no sampling. Four tools came back from tools/list, and each carries a complete annotation block (readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false), which is what lets a client decide without asking a human whether a call is safe.
com.clickdecay/mcp-server 1.0.0 answers in the official registry as active, published 2026-11-12T23:10:41Z, with the remote declared as the same URL and the description CTR change, click decomposition and known Search Console data anomalies for numbers you provide.
The server's own instructions string states the contract before any tool is called: it calculates CTR changes from Search Console totals that the user provides, and every number carries an evidence label. That is the whole design, and it is the reason this entry exists.
What the four tools do¶
compare_ctr_periods two equal-length, non-overlapping periods of
clicks and impressions -> CTR for each, the click
delta, and the split of that delta into an
impressions component and a CTR component
project_clicks_at_ctr impressions x a CTR you choose (a what-if)
impressions_needed_for_clicks clicks / a CTR you choose (sizing a target)
check_gsc_reporting_issues the known Search Console notices overlapping
your dates, with Google's own source link
The calls that were driven¶
Eleven calls against the endpoint. The comparison below used illustrative totals, not any site's data: a prior period of 3,000 clicks from 100,000 impressions and a current period of 3,150 clicks from 125,000 impressions, 28 days each and non-overlapping.
priorCtr 0.03 observed
currentCtr 0.0252 observed
clickDelta +150 observed (the business already knew this)
impressionDelta +25,000 observed
ctrChangePp -0.48 observed percentage points
relativeCtrChangePercent -16.0 observed
clicksAtPriorCtr 3,750 modelled what 125,000 impressions earn at 3%
gapToPriorCtrScenario -600 modelled clicks given up to the CTR change
impressionsComponent +750 calculated clicks gained from more impressions
ctrComponent -600 calculated clicks lost to the CTR change
warnings ["NO_AI_CAUSATION"]
versions server 1.0.0, engine 0.1.0, contract 0.2.0,
anomalyRegistry 2026-10-12
The control is the point: the decomposition was re-done by hand and matched. 125,000 impressions at the prior CTR of 3 percent is 3,750 clicks; 3,750 minus the observed 3,000 is the +750 the server calls the impressions component; 3,150 minus 3,750 is the -600 CTR component; and the two add back to the observed +150. A reporting tool that gets this wrong is worse than no tool, and this one is reproducible to the click on the arithmetic it declares calculated.
The two what-if tools answered the same way, and both said what they are: project_clicks_at_ctr with 125,000 impressions at 2.52 percent returned 3,150 labelled modelled; impressions_needed_for_clicks for 4,000 clicks at 2.5 percent returned 160,000, also modelled.
The behaviour that earns the guide¶
Google published an impressions-logging anomaly running 2025-05-13 to 2026-04-27 that makes CTR reading unreliable for any site in that span. Ask Click Decay for a comparison that touches it and the answer changes shape rather than changing tone:
periods 2026-02-01..02-28 vs 2026-03-02..03-29 (28 days each)
priorCtr, currentCtr, impressionDelta, ctrChangePp,
relativeCtrChangePercent, clicksAtPriorCtr,
gapToPriorCtrScenario, impressionsComponent, ctrComponent
value null, nullReason "GSC_REPORTING_ANOMALY" (each one)
clickDelta +200 observed <- clicks are still compared
anomalies [{ id: GSC_IMPRESSIONS_ANOMALY_2025_2026, ... }]
Every figure that depends on impressions is withheld with a reason code, the anomaly is named in the response, and the one thing that can still be trusted - the click count - is still returned. Run the same call on dates after the anomaly and the numbers come back. That is a calculator that will not produce a number it cannot stand behind, which is a different thing from a calculator that produces a number with a disclaimer underneath.
check_gsc_reporting_issues with no arguments returns the registry the server is working from, two records at reading: GSC_IMPRESSIONS_ANOMALY_2025_2026 (affects impressions, ctr and position; applies to all; effect block_ctr_diagnostics) and GSC_BULK_EXPORT_MISSING_DAYS_2026 (2026-02-28 to 2026-03-01; affects rows; applies to bulk export; effect warn). Both carry sourceUrl https://support.google.com/webmasters/answer/6211453 - Google's own notice - rather than a vendor claim.
The refusals, and one defect¶
Three bad inputs were tried and all three were refused with a machine-readable code and isError: true:
28 days vs 14 days {"errors": [{"code": "PERIOD_LENGTH_MISMATCH", "field": "periods"}]}
periods that overlap {"errors": [{"code": "PERIOD_OVERLAP", "field": "periods"}]}
300,000 clicks on 100,000 impressions
{"errors": [{"code": "CLICKS_GT_IMPRESSIONS", "field": "prior.clicks"}]}
The defect, measured and recorded: those refusals arrive as HTTP 200 with structuredContent.errors and no content block at all. A client that renders only content - which is most of them - shows the operator an empty answer rather than "the two periods must be the same length". The same shape appears in the null path: ctrPercent: 0 on impressions_needed_for_clicks returned {"value": null, "nullReason": "ZERO_RATE"} with isError absent, so a null is not an error, it is a value with a reason attached - correct in the data model, easy to drop in a chat window.
A second, smaller artifact: two observed figures come back unrounded, ctrChangePp as -0.47999999999999987 and relativeCtrChangePercent as -15.999999999999993. A formatter in the client hides it; nothing in the response does.
What it does not do¶
No Search Console connection, no OAuth, no file upload, and no reading of your property - the totals are typed or pasted, which is either a limitation or the entire point, depending on whether you are willing to hand a third party a Search Console scope to get arithmetic. The vendor states the tool takes numbers, dates and the name of the data source only, calculates in memory, stores nothing, and publishes a roughly 60 requests a minute limit per connection. It also does not explain why clicks or CTR moved, and every comparison says so in the NO_AI_CAUSATION warning rather than leaving the reader to assume the assistant concluded something.
How to connect¶
Claude Code claude mcp add --transport http clickdecay https://clickdecay.com/mcp
Claude web/app Settings -> Connectors -> add a custom connector named Click Decay,
address https://clickdecay.com/mcp, no authentication
ChatGPT developer mode in the advanced connector settings, then a connector
with the same address and no authentication
Cursor {"mcpServers": {"clickdecay": {"url": "https://clickdecay.com/mcp"}}} in .cursor/mcp.json
VS Code {"servers": {"clickdecay": {"type": "http", "url": "https://clickdecay.com/mcp"}}} in .vscode/mcp.json
If you run several data surfaces in one assistant, the practical order is: read the totals from your Search Console connector, hand the two period totals to Click Decay, and keep its labelled answer beside the reader's raw rows. The decomposition is arithmetic on published numbers, so the labelled calculation is the part you can quote to someone who will check it.
Where it sits in this catalog¶
The SEO slice here is well covered by readers and connectors - Google Search Console (ni-c), Otto MCP, AgentGrown MCP, Metrinel, GSCPAP, Sequel, SEOmatic, Sorank, HyperGSC and SearchDock all get the numbers out of a property. Click Decay does not read a property at all: it takes the totals and adjudicates the arithmetic, keeps the observed apart from the modelled, and withholds what a Google notice makes untrustworthy. That refusal is the reason it is catalogued and the readers are not duplicated.
What does the evidence label on each number actually mean?¶
observed is arithmetic on the numbers you supplied. calculated is an exact decomposition of the change - the impressions component and the CTR component add back to the observed click delta. modelled depends on an assumption you or the tool chose, such as a target CTR. not_established, and the nullReason codes beside a null, mean the server declined to produce a figure and said why.
What happens when a known Search Console issue covers the dates?¶
Every impressions-dependent figure comes back null with nullReason: "GSC_REPORTING_ANOMALY", the issue is listed in anomalies with Google's source link, and the click comparison is still returned. The tool does not quietly compute a CTR it knows Google's own data makes unreliable.
Does it need Search Console access or an API key?¶
No. There is no account, no key, no OAuth and no file input. You supply totals and dates; the server computes in memory and stores nothing. That also means it cannot verify the numbers you give it - the labels describe the arithmetic, not the provenance of the inputs.
What are the limits worth knowing before you rely on it?¶
Roughly 60 requests a minute per connection, four tools, no per-query, per-page or per-property analysis, and refusals that arrive without a human sentence - a client that shows only content will look empty on a bad input. Two observed figures come back unrounded, so format them in the client.
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.