How a home gets flagged.
FirstFlag reads public records and prints what they hold. This page is the whole test: the sources, the three signals, the order homes come back in, the thresholds CQC publishes and how closely our own data tracks them, and the parts of the claim we cannot yet stand behind. It is written so you can take any flag we show you, open the home's own CQC page and check us.
Regulatory data: Care Quality Commission, Open Government Licence v3 · Company data: Companies House · Enrichment pass last run 30 July 2026
Counted from the enrichment cache this page is built from, not quoted from memory.
What we read, and what each source can actually tell us.
Five public sources, and nothing else. None of them is a purchased list, a scraped opinion or an inferred contact. Where a record is silent, the product is silent: a field the source does not hold renders as absent rather than as a plausible guess.
One row per registered location, carrying the overall rating, the five domain ratings (Safe, Effective, Caring, Responsive, Well-led) and the date CQC published each one. That is all it carries. No registered manager, no directors, no phone numbers, no bed counts, no inspector commentary and no coordinates. Only one of the 22 values in the register's location-type field means care home; the rest are home-care agencies, GP practices, dentists and similar, and counting those as care homes would overstate a scan by roughly three times.
The location and provider records: the registered manager or the recorded absence of one, beds, address, registration date, service type, specialisms, the full rating history, and under the single assessment framework the percentage score for each key question plus the inspector's own evidence text. Fetched ahead of a partner opening their portal and committed as a file, so the page needs no API key, adds no request latency and cannot be degraded by a CQC outage. The trade is freshness, which is what the refresh cadence below governs.
The published latest ratings download, one row per active location per domain, used for national and regional counts. The data vintage is taken from CQC's own file date rather than from our clock, so re-running the refresh against an unchanged file produces an unchanged page.
CQC's monthly register of active locations, the same download page as the ratings file. It names the provider's nominated individual, the person the provider has told CQC is responsible for supervising the regulated activity, and the registered manager for each location. The nominated individual is the reason this record is worth reading over a bare directors list: a company's officers include people who have never been near a home, while the nominated individual is the one the provider itself has put forward to the regulator. Where that person is also a Companies House director, the record says so rather than listing them twice. Vintage is taken from the file's own "as at" date, not from our clock.
Read on the company number CQC itself publishes against the provider, so there is no fuzzy name match and no guessed join. It returns the named directors and their appointment dates, the incorporation date, whether the annual accounts and confirmation statement are filed on time, and any insolvency history. Where CQC's provider name and the name on that company number have drifted apart, the dossier says so rather than picking one.
Two tiers, and the difference is worth stating. The preferred source is the contact panel of an NHS.uk page whose final URL still carries that home's own CQC location id, which ties the address to that registration and no other: 140 of the 735 flagged homes are sourced that way. Where NHS.uk holds no address, the pass reads the contact page of the home's own site, and where no site is recorded against the home, the operator's. Only addresses on that site's own domain are kept. That is a weaker tie, because an operator publishes for a group rather than for one home, and it accounts for the other 119. Both counts are taken after that second pass is cleaned up: malformed addresses, corporate role inboxes (press, HR, recruitment, fundraising, volunteering) and any address naming a different home are dropped. So these are the homes a dossier will actually print an address for, not the homes we happen to hold a string against. What survives is never assembled from a pattern. We print the address an operator has published and never a guessed firstname.lastname@, which is why the remaining 476 flagged homes carry no email at all.
A signal is a condition on a published field.
Not a model output, not a sentiment reading, not an opinion about a home. A signal is either true of the public record or it is not, and every one of them can be checked against CQC's own page for that home in about a minute. There are three of them. A home is flagged when it trips at least one, and the sharpest is the only one the register on its own can support:
The overall rating is Good or Outstanding, and Well-led has already been rated Requires improvement or Inadequate.
Leadership is rated down before the headline rating catches up, and the headline is what everybody else is reading. The search box on cqc.org.uk cannot run that query, though the monthly ratings file can be made to, which is the subject of the section below. A patch derived live from the register is built on that test alone, so every home in one carries it. The enrichment pass then adds the other two codes where the record supports them: the first stands as a signal in its own right, the second only ever sharpens a home that is already flagged.
Manager gone
CQC records no registered manager against the home's regulated activity: the post is unfilled, and nobody currently holds the accountability the register exists to place. The search box on cqc.org.uk cannot filter on it, but CQC's monthly care directory with filters has a column for it, so the bare list costs an afternoon and nothing else. What it does not come with is any of the six things set out below. True of 91 of the 735 flagged homes held today.
Well-led cracking
The test above: the overall rating still reads Good or Outstanding while Well-led has already been rated Requires improvement or Inadequate. The sharpest signal the register alone can support, and the one every home in a register-built patch carries.
Long since inspected
Six years or more since a published inspection on a home still rated Good or Outstanding. On its own this is the CQC backlog rather than a signal, and a long gap is ordinary across homes CQC still rates Good. P3 ranks a flagged home. It never flags one.
Signals overlap. One home can carry all three, so the counts do not sum to the number of flagged homes.
You can build the flag list yourself. Here is the file.
CQC publishes its whole ratings table as a monthly spreadsheet, the care directory with ratings, one row per location per domain with Well-led among them. Open it, filter to care homes in England, pivot the domain column against the location id, and keep the rows where overall reads Good or Outstanding while Well-led reads Requires improvement or Inadequate. That is the P2 list. It costs nothing and it takes about an hour.
The manager signal is no different, and it is easier. The companion file, the care directory with filters, carries a Registered manager column beside a Care home? column. Filter to care homes, keep the blank cells, and that is the P1 list. One AutoFilter, no join, no pivot.
We are telling you this on our own pricing argument because you would have found it, and because the version of this page that hid it was not worth reading. The regulator's search box genuinely cannot run that query, which is a real and narrow fact about the website. It is not a fact about the record, and selling it as one would be the same move we flag operators for.
What the spreadsheet gives you is a column of location ids. Here is what it does not give you, which is the entire product:
If the list on its own is what you need, take the file. Nothing here is under licence and we would rather you spent the afternoon than the money. What is being charged for is items one to six, and they are the ones you can price against your own hourly rate.
What the score is, and why the portal does not lead with it.
Every flagged home carries a score from 0 to 99, computed from register fields only. It is deterministic: the same record produces the same number on every render, with no random element and no weighting learned from data we do not have.
The score is a ranking aid, not a headline, and it is not a probability of anything. On register fields alone a large part of any patch scores identically, and a column of repeated numbers would claim a precision the data does not have. So the portal sorts on the date CQC published the Well-led rating and shows that date in place of a score. Ties fall to the score, then to the staler inspection, then to the CQC location id, which is why two renders of the same patch come back in the same order.
Where the Good line actually falls.
CQC scores each quality statement from 1 to 4 and divides the total by the points those statements could have scored, so a key question is published as a percentage rather than as a mark out of a hundred. It publishes where that percentage becomes a rating: 88 per cent and above is Outstanding, 63 to 87 is Good, 39 to 62 is Requires improvement, and 38 or lower is Inadequate. Those are CQC's figures, in its guidance on how it reaches a rating and again on its public ratings and scores page.
One published rule overrides all four bands, and it is the part worth a consultant's attention. Where a key question scores in the Good range but any single quality statement scored 1, CQC holds the rating at Requires improvement. Where it scores in the Outstanding range but a statement scored 1 or 2, the rating is held at Good. Clearing 63 is necessary and not sufficient, and a home can be marked down on a domain its own total says it passed.
Across every scored key question at every current assessment we hold, 742 judgements, 706 sit in the band their published rating implies. That is 95.1 per cent. The rest are three known things rather than a fuzzy line: 33 sit exactly one point under the floor for their own rating, which is what a rounded percentage does; 0 carry a score of zero beside a rating well above Inadequate, which is a null stored as a zero on our side and is why every derivation drops a score of zero or less; and 3 score inside a higher band than the rating they were given, which is the limiter above doing exactly what CQC says it does. Nothing is left over.
The distance between a home's score and 63 is the size of the shortfall in CQC's own units and nothing more. It does not promise that reaching 63 next time produces a Good, because the limiter can hold a home under the band it has reached, and it does not forecast an inspection. Nothing on this site does.
The marks underneath the score, and what they determine.
CQC scores each quality statement from 1 to 4 in its own words: 1 is evidence of significant shortfalls, 2 some shortfalls, 3 a good standard, 4 an exceptional one. Those marks are published against every key question, and the arithmetic on top of them is CQC's own, from the same published guidance as the bands above: the percentage is the marks' mean, over 4, and the rating is the band that percentage falls in, unless a statement scored 1, in which case the limiter applies.
So the marks, plus that published arithmetic, should be enough on their own. We threw away CQC's percentage and CQC's rating, kept only the marks, and rebuilt both from the published rules. Across the 643 key questions CQC assessed at a current assessment in our cache, the rebuilt rating matches the one CQC published on 640 of them, 99.5 per cent. The 3 it misses are all rated one band below what the marks give, which is the professional judgement CQC says it applies on top of the principles, running in the cautious direction. A model that predicted those away would be wrong about the regulator.
The published percentage is a whole number and the marks behind it usually are not. A very common shape of marks averages exactly 62.5, which CQC prints sometimes as 62 and sometimes as 63 while rating it Good either way: 31 of the 743 key questions we hold marks for sit on that edge. So a home reading 62 has not necessarily been marked down, and anyone selling on “one point from Requires improvement” is selling a rounding artefact.
The limiter moves the printed figure as well as the rating, which CQC does not spell out. 46 key questions in the cache have marks averaging above 62 and print at exactly 62, because one statement scored 1. On those, the number CQC prints understates its own marks, and the gap between the two is the only thing a consultant needs to open a conversation.
2 of the 743 print a percentage we cannot rebuild from the marks CQC published beside it, and we have not worked out why. They are counted here rather than dropped.
How often this moves, and what is stale on purpose.
The scan is built from CQC's published register and refreshed as that register updates. A patch is stamped with the date of the last successful ingest run and never with the clock, so the figures on a portal are dated to the day they were true rather than to the moment you loaded the page.
The deep enrichment pass, which is what supplies managers, directors, filing status, inspector evidence and contact routes, runs separately and is committed as a file. It last ran on 30 July 2026. That is a deliberate trade: no API key in production, no request latency on a partner's page and no dependency on CQC being up at the moment a prospect opens a link, in exchange for the enrichment being exactly as old as the last pass.
New flags appear as they trip. A home whose Well-led rating was published in the last 24 months is counted separately as the live end of the pool, because a home working through a finding from last spring is a different conversation from one marked down years ago and never revisited.
What this cannot tell you.
The strongest thing we can do on a page like this is be exact about where the method stops. Every item below is a real constraint, not a disclaimer.
We cannot show you a hit rate.
FirstFlag has no customers, so there is no customer result and no accuracy rate from live use anywhere on this site. What exists is a backtest: every flagged home in four years of CQC's ratings files, followed to July 2026 and published in full on the evidence page, including the findings that count against the flag. A backtest is history, not a promise about the next inspection. When a figure measured on FirstFlag's own flags in live use exists, it will appear here.
A flag is not a prediction that a home will fail.
It says a published record shows a defined condition now. Homes recover, operators fix things, and some flagged homes will be rated Good at the next assessment. What the flag buys you is the order you find out in.
Public records lag reality.
A rating is published some time after the visit that produced it. A registered manager can be appointed the day after we read the register. Nothing here is real time, and nothing here should be treated as certainly current on the morning you pick up the phone.
We date a finding only where CQC dates it.
There is a date field on an assessment record and it is not the date of the visit. We could not establish a reliable relationship between the two, so we do not infer one. Where CQC prints the assessment window in its own report we carry that window word for word, and where it does not, the finding carries no date at all.
A missing field stays missing.
If the record holds no manager, no director and no email, the product prints nothing in that space. Absence in a dossier means absence in the public record, not absence in the world, and a register-derived patch carries ratings and dates only.
England only.
The register we read is CQC's, which covers England across 9 regions. Wales, Scotland and Northern Ireland have their own regulators and are out of scope.
Decision support, not advice.
Signals are derived from public records that can lag reality, and we do not warrant that any flagged home will fail, improve or engage you. That sentence is also in the terms, and it is the same claim this page makes.
Every dossier links back to the record it was built from.
Each one carries a link to that home's own page on cqc.org.uk. Open it, read the domain ratings and the publication dates, and compare them to what we printed. If the register says something we did not, the register is right and we want to hear about it.
A territory patch is built once and then kept refreshed as the CQC register updates, and it is quoted against the patch you need rather than off a list. A territory is drawn round the area a firm actually covers, so it can span more than one CQC region.