We Ran the Same Content Audit Across 24 of Our Own Sites in One Day
- RedClaw Performance Team
- RedClaw Performance Team
- Date
- Reading time
- 17 min read
TL;DR: on 8 September 2026 we pointed one content-acceptance script at 24 sites we own, in a single day, with the same thresholds and the same gate definitions for all of them. Across the ten with complete counts, 340 of 478 pages pass after fixes (71%). That number is not the finding. The finding is that three sites opened at 0/61, 0/42 and 0/44: completely red, and only two of those three were configuration. The fourth one's 44/44 was real. Distinguishing the two is the most portable thing we got out of the day, and it is a two-step rule, not a one-step rule. This piece covers the eight dimensions, what the distribution across 24 sites actually looked like, and how to run the same thing on your own portfolio.
Why audit yourself instead of launching another site
The motivation determines whether this is worth a day of your time, so it goes first.
At any moment there are two places to put production capacity: a new site, or the sites you already have. New sites are intuitively attractive because they have no history.
Our own data rejects the intuition flatly. Our main site draws 110,217 impressions in 90 days. One site with genuinely good content, stuck at average position 31.4, draws 31,884. The marginal cost of moving either of those up one position is far below 12 to 18 months from zero, and we have four new sites' worth of evidence of what those 12 to 18 months look like: 360 pages, 48 clicks.
So the audit is not a scoring exercise. It answers a capital-allocation question: of the pages that already exist, which ones move if touched, and which are permanent dead weight?
Answering that requires one instrument. Measure different sites with different standards and the results are not comparable, which means they cannot inform an allocation decision.
The instrument
Eight dimensions. The design decision that matters is not the list. It is which of them are gates (fail and the score is zero) and which are deductions:
| # | Dimension | Type | Checks |
|---|---|---|---|
| 0 | Indexing worth | Human gate | Remove every cited source. What is left? |
| 1 | Article quality | Deduction | Length, headings, summary, description length |
| 2 | AI fingerprints | Deduction | Stock phrasing, paragraph-rhythm variance, opener/closer shape, em-dash density |
| 3 | Original insight | Deduction | First-hand markers; reproducibility by any model |
| 4 | GEO readiness | Deduction | Citability; extractable conclusions |
| 5 | Authoritative sourcing | Gate | Fewer than 3 distinct external domains fails outright |
| 6 | Internal links | Deduction | Count, trailing slashes, link to a money page |
| 7 | Release discipline | Deduction | Publication date, scheduling, factual currency |
Making "fewer than three external domains" a gate rather than a deduction is the most contested and most effective choice in the whole instrument. The reasoning: an article with no external sources usually has no thesis. It is a topic re-typeset. A deduction lets it recover elsewhere; a gate does not.
It is also, empirically, the single largest systematic gap across the portfolio, by a wide margin, as the next section shows.
Gate 0 is human, and it comes before everything
"Indexing worth" is not in the script, because it cannot be. Its test is one sentence:
Strip out every source you cited. What is left?
If the answer is nothing, kill the topic. This comes straight from Google's own language: the spam policies↗ name aggregator sites that add no context beyond scraped source data, and the 2023 guidance on AI-generated content↗ settled the criterion years ago. Production method is irrelevant; usefulness is not.
Certain formats are structurally at risk here: rankings, roundups, glossary entries, summaries of official announcements, competitor comparisons. None are forbidden, but shipping one requires attaching first-hand data or an explicit judgement, or it is indistinguishable from the other fifty pages built on the same public dataset.
Gate 0's second question: does anybody search this page?
After the August 2026 spam update we added a second question to Gate 0: is there search demand for this page at all?
That update's criterion was exactly this, and it operates at site level: mass-producing pages nobody searches for causes the whole domain to be re-evaluated, including pages that were fine. Our measurement of 20 of our own properties across that rollout is written up separately alongside the tooling findings.
Answer this with real impression data, not tool estimates. A tool gives you a model's output; the Search Console Page Indexing report↗ and your own impression numbers give you what actually happened on your domain.
What 24 sites looked like
The distribution is bimodal
Sites with complete counts; client assets excluded:
| Site | Pages | Pass (post-fix) | Rate |
|---|---|---|---|
| A (regulatory tools) | 77 | 77 | 100% |
| B (lottery tools) | 69 | 69 | 100% |
| C (analytics tools, EN) | 23 | 23 | 100% |
| D (Philippines vertical, EN) | 12 | 12 | 100% |
| E (odds tools) | 74 | 62 | 84% |
| F (AI tools, EN) | 61 | 48 | 79% |
| G (holiday tools) | 33 | 26 | 79% |
| H (lifestyle content) | 37 | 13 | 35% |
| I (pet content) | 50 | 3 | 6% |
| J (affiliate content) | 42 | 7 | 17% |
| Total | 478 | 340 | 71% |
Read 71% carefully. It is post-fix, and "fix" meant very different things per site. Some only got a corrected config file; some had external sources added to dozens of articles. It is not a benchmark you can compare yourself against. It is our own starting line.
The informative part is the shape: four sites at 100%, two at 79–84%, then a cliff to 35%, 17% and 6%. There is almost nothing in the middle.
That absence tells you something operational. Passing is not a per-article random variable; it is a per-site process difference. A site whose editorial workflow has "every article needs external sources" baked in lands near 100%. A site without it lands near zero. You do not get a site that happens to be half-and-half.
Which means: fix the process, not the articles. Site I's 3 out of 50 is not 47 independent quality problems. It is one workflow gap occurring 47 times.
The most common failure, by a distance
Counting every gate failure across the portfolio:
- Fewer than 3 external source domains: first by a wide margin
- Missing summary / TL;DR
- Internal links missing trailing slashes
- Body length below threshold
Site I had 30 of 50 pages citing zero external domains, not "fewer than three", zero. Site J had 27 of 42 failing the summary check. Every one of site H's 37 pages was under the length threshold, averaging 1,413 characters against a 3,000 minimum.
Why is sourcing first? Because it is the only rule that requires leaving the editor. Length is solved by writing, headings by formatting, a summary by adding a paragraph. Only "find three distinct authoritative domains and confirm they still resolve" requires going somewhere else: and any step that requires leaving the editor gets skipped unless it is in the workflow.
The most portable rule: what an 80% hit rate means
This is the highest-value output of the whole exercise, and three sites forced it out of us.
| Site | First run | The gate that fired |
|---|---|---|
| F | 0 pass / 61 fail | Two rules, at 60/60 and 61/61 |
| J | 0 pass / 42 fail | No config file in the project root |
| K, ad tutorials (EN + local) | 0 pass / 44 fail | One rule, 44/44 |
Site F's two failures:
- 60 of 60 reported "internal links missing trailing slash." That site's canonicals have no trailing slash, the config never declared it, and the script defaults to expecting one. Following the tool would have meant hand-building 60 redirects.
- 61 of 61 reported "no link to a money page." The config listed
["/tools/", "/try"]with leading slashes; joined into a regex that becomes^/(/tools/|/try…), which never matches. Corrected, the number genuinely missing a money-page link was five.
Site J was simpler: no config file at all, so the script looked for default field names that site does not use.
Hence the rule:
When a single gate fires on more than 80% of files, assume configuration first.
Because content defects are scattered. Fifty articles written by different people in different months fail in different ways: one lacks sources, one is short, one has no internal links. They do not line up on one rule. A 60/60 is the shape of a setting.
And the counter-example that keeps it honest
Site K matched that shape exactly: 44/44 on one gate, zero external source domains, and it was real. All 44 tutorials had no outbound link anywhere; grep https?:// returned nothing.
So the rule ships in two steps, and the second is not optional:
- One gate over 80%? Go read the config. Do not touch articles.
- Config confirmed correct? Then that 80% is a genuine site-level process gap: fix it once, across all 44.
Step one is now in the script: any batch run where a single gate exceeds 80% prints a line saying it looks like configuration, with the matching hint. It does not decide anything. It makes sure the signal is not missed.
A fourth site was a variant of the same illness: its content lives in React component props, the extractor could not read them, and 12 gates went red. Fixing the extractor took it to zero. Also not an article problem.
Per-article passing is not per-batch passing
There is a class of problem that per-article checking cannot see by construction: every article passes, and they are all the same shape.
Our batch check computes three variances: body-length coefficient of variation, heading-count CV, and the distribution of opener and closer types. Too little variance is a red flag. The four failed sites were red on all three, 154/154, 74/74, 78/78 on a single skeleton.
Variance matters because content from one mould is predictable to a reader and identifiable to a machine. And an identifiable pattern judged low-value is judged as a batch, not article by article. That is precisely how the August 2026 update worked: site-level.
No per-article score protects against this, because the problem is not located in any article.
Run this on your own portfolio
You do not need our script. The sequence is what transfers:
1. Write the config before the first run. At minimum declare three things: whether your site uses trailing slashes, what your frontmatter fields are called, and which paths count as money pages. Skip this and your first run is 0/N, and you will spend a day editing articles that were fine.
2. Do not read scores on the first run. Read the distribution. Compute the hit rate per gate. Anything over 80%? Return to step one.
3. Once config is confirmed, split findings into site-level and article-level. High hit rates are process problems, fixed in one pass. Scattered ones are the per-article queue.
4. Verify that sources resolve. Do not just count domains. Request each one. Three classes: alive (2xx/3xx), blocked (401, 403, 405, 429, 451, 503: bot walls, flag for a human), dead (404, 410, DNS failure, timeout). Only the third is a gate. And retry a 404 with GET before believing it: some servers, including Google's own help pages, answer HEAD with 404 and GET with 200.
5. Re-run a rendered-layer check after deploying. That layer does not exist before deployment, so your CI cannot reach it. This is the step almost everyone skips. Ours is open source↗.
6. Look at batch variance alongside the scores. Length, heading count, opener shape. All identical means the template needs redesigning regardless of how well individual articles score.
After the audit: improve or prune
The report produces exactly two decisions per page: add to it, or remove it. Our tree:
Any clicks in 90 days?
├─ Yes ────────→ Never prune. Improve it, and use it as an internal hub.
└─ No
├─ Any impressions?
│ ├─ Yes (pos < 20) → Improve: title, snippet, internal links. It is within reach.
│ └─ Yes (pos ≥ 20) → Ask whether this domain can reach that term at all.
│ If not, it is not a content problem. Stop adding content.
└─ None at all ─→ Live for 12 weeks yet?
├─ No → Wait.
└─ Yes → Prune candidate.
Four exceptions that are never pruned, each learned the hard way:
- Any page with any clicks in 90 days, even one. A click is a verified human need; deleting it throws that away.
- Anything under 12 weeks old. New pages are still being evaluated; deciding on that data is deciding on noise.
- Internal hubs. Pruning a page many others point at breaks all of those links.
- Pages carrying first-hand data. That data will not exist a second time. Poor position is a keyword-selection problem; deleting the page throws out the evidence with it.
Always check Search Console before pruning. We have made this mistake: deleted a batch of pages that "looked useless", and one of them was still ranking at position 6.7, simultaneously ranking and sending people to a 404. What a page ranks for and what you assume about it are different things.
How often
Our cadence: weekly per site, quarterly across the portfolio.
The weekly pass stops debt accumulating. The quarterly cross-section catches what is only visible in comparison: false negatives, template repetition, workflow gaps. None of the three tooling blind spots we found that day would have surfaced in a single-site report.
If you only ever do one, do the cross-section, because it tells you which sites deserve the weekly one.
The tooling half of this audit is in three things your SEO checker cannot see; the four-site post-mortem is in 360 pages, 48 clicks; the vertical-specific version is our complete iGaming SEO guide.
If you would rather have somebody run all of this on your sites weekly and hand you the numbers, that is our SEO site build and management service.
FAQ
How long does auditing 24 sites take?
The script is fast: minutes per site. The time goes into per-site configuration and the human Gate 0 call. We produced 24 reports in one day working in parallel, but every report still needed a person to read it. The script produces a score; Gate 0 is a judgement.
Is 71% passing good?
There is no external benchmark, so the number means nothing to anybody but us. Its value is as a starting line: 340 pages hold up, 138 need work, and we know how many of those 138 are workflow gaps rather than individual failures. The only useful comparison for your first cross-section is your second one.
Why gate on fewer than three external domains rather than deduct?
It is the single best discriminator between "has a thesis" and "re-typeset a topic", and it is the only rule that requires leaving the editor. As a deduction it gets compensated elsewhere and then nobody ever does it. Empirically it is also the most common gate failure across the portfolio.
Does one gate firing everywhere always mean a config problem?
No, and this is the sentence to be careful with. Two of our three fully-red sites were configuration; the third's 44/44 was real, with no outbound links anywhere in 44 tutorials. The rule is check config first, not "ignore it." Once config is clean, that 80% is a genuine site-level gap.
What is your pruning threshold?
12 weeks live, plus zero impressions, plus not in the four exceptions (has clicks, under 12 weeks, internal hub, carries first-hand data). And always check actual Search Console positions first, we once deleted a page still ranking at 6.7.
I do not have 24 sites. Is this still useful?
Yes, and it is less work. The single-site version is: write the config, run once, check whether any gate exceeds 80%, verify source liveness, and re-run a rendered check after deploying. The cross-section's specific value is catching false negatives, which needs multiple sites: but those five steps run on one site and already cover most of what you will actually hit.
Related Posts
- strategyThree Things Your SEO Checker Cannot See (And the Tool We Built for One of Them)One checker, 24 sites, one day. Three opened at 0/42, 0/61 and 0/78 on configuration alone, and one cited three broken sources across 92 articles with the gate green.
- strategy360 Pages, 48 Clicks: An SEO Case Study of Four Sites That FailedFour of our own sites: 360 pages, 33,615 impressions, 48 clicks in 90 days. The page with the most clicks has no first-hand data at all; it just ranks 7.6.
- strategyWhite Label CMS for Agent and Franchise Networks: A Buyer's ChecklistA white label CMS checklist for head offices with agents or franchisees: who owns the domains and data, what agents can touch, what the price covers, and how you leave.