Digimetrics Platform
QA & Fix Report
A hands-on audit of all 36 tools on platform.digimetrics.ai, tested primarily against Hermès USA — a large, JS-rendered, globally-indexed enterprise site — and cross-checked against a small local SME (Furnituresg.com.sg) to see what breaks at each end of the spectrum.
How to read this report. Severity: Critical breaks or actively misleads · Friction confusing, inconsistent, or inconsistent-but-working · Note wording/copy/suggestion. Every finding below includes exact reproduction steps instead of a screenshot — this version of the report was built without image-embedding support, so each claim is written to be independently re-run and verified rather than eyeballed. Where a tool made a checkable factual claim (robots.txt, structured data, H1 tags, meta description), it was independently verified against the live page via direct DOM inspection, not taken on faith — marked with a ✓ ground-truth verified tag below.
01 Executive summary
The underlying product is genuinely strong. Every tool tested — across two very different sites — produced real, mostly well-reasoned, business-specific output, not generic filler. Several tools (Competitors Identifier, On-Page Optimisation, AI Mentions Tracker, Backlinks Explorer, Persona Generator) are legitimately best-in-class relative to standalone competitors charging $99–500/month each for narrower capability. The problems found are not "the AI is bad" — they're integration and trust issues: tools disagreeing with each other and with reality, implementation details leaking into user-facing copy, and a couple of rough edges (fabricated-looking numbers, competitor name-drops) that would rattle a skeptical first-time user even though the analysis underneath is sound.
The Hermès test specifically surfaced the report's single most important finding: on a large, modern, JavaScript-rendered site, three tools produced confidently wrong pass/fail verdicts about the site's own SEO fundamentals — not just disagreeing with each other, but verifiably wrong against the live page. That's a materially bigger problem than a UX inconsistency, because it's exactly the kind of site — funded startups, enterprise brands, agencies pitching bigger clients — Digimetrics needs to win.
02 Critical findings
The highest-leverage fixes — each affects multiple tools or carries real trust/compliance risk. Ranked by impact.
1. Crawl-based tools produce confidently wrong verdicts on modern, JS-rendered sites
Critical · Highest priorityThis is the single most important finding in the report. Three tools — AI Readiness Check, Full Site Audit, and llms.txt Generator — were run against hermes.com/us/en within minutes of each other. Every one of them produced at least one confidently wrong "Fix" verdict, and no two tools fully agree with each other.
| Check | AI Readiness Check | Full Site Audit | llms.txt Generator | Ground truth (live DOM) |
|---|---|---|---|---|
| robots.txt exists | ✗ Fix (missing) | ✗ Does not exist | ✓ Pass — correct | Exists — 150+ lines, 36 sitemaps |
| Structured data (JSON-LD) | ✗ Fix / No | ✓ Ready / Yes — correct | not checked | Exists — Organization schema present |
| Semantic HTML | ✗ Fix | ✗ Gap | not checked | Exists — header, main, 3× nav, footer |
| Meta description | ✗ Fix | ✗ Missing | ✗ No meta description | Exists, substantive copy |
| H1 tag present | — | — | — | Missing — 0 found, genuine gap |
| Pages linked from home | not checked | not checked | "0 pages found... empty site" | False — extensive category nav |
This session's re-test adds a calibration point. Re-running Technical SEO Crawler against Hermès and independently checking document.querySelectorAll('h1') on the live page confirmed 0 h1 tags — the crawler's claim was correct this time. The pattern across both sessions: simple, static checks (robots.txt fetch, H1 presence, third-party backlink data) come back accurate; anything requiring the tool to parse the rendered page — meta tags, JSON-LD, semantic landmarks, internal links — comes back wrong, and specifically wrong in the direction of "nothing found."
The crawler's fetcher likely doesn't execute JavaScript, or Hermès' infrastructure serves reduced content to non-browser user agents. Either explains every discrepancy above: static-file checks work, DOM-dependent checks fail toward "not found."
Why this outranks the contradiction finding below: Furnituresg.com.sg is small and presumably server-rendered — the tools were likely built and tested against exactly that profile, which is why they work there. Hermès represents the tier of customer (funded startups, enterprise brands, agencies) built on modern JS frameworks. If the crawler silently fails on that entire tier and reports specific, confident "you're missing X" findings instead of "we couldn't fully read this page," a prospect's first experience with the product can be a wrong, embarrassing report about their own site.
2. AI Image Generator and AI Video Generator both fail on all plain text-to-generate requests
Critical · Blocks core usageConfirmed four separate times across two sites and two sessions — this is not an edge case, it looks like both tools may not currently work at all for basic generation.
- Open AI Image Generator or AI Video Generator, choose "Guided — answer a few questions"
- Enter any plain-text description (no image reference, no re-edit) — e.g. "An orange Hermès Birkin bag on a marble pedestal, studio lighting"
- Click Run tool
Result, verbatim (this session, Image Generator): "Your source image didn't upload correctly or isn't recognized by this tool — it needs to be an image created within this system to work properly." Followed by a "Technical detail" line on the Video Generator's failure this session: "That source image could not be read. It must be an image generated here."
Sharper root cause, found this session: no source image was ever provided in either request — these were plain guided text-to-generate submissions. Both tools throw the exact same "source image" error regardless, which strongly suggests one shared backend code path unconditionally checks for a source image parameter even when the request mode never collects one. This is precise enough to hand to engineering directly: check whatever validation step runs before generation on both tools for an unconditional source-image read.
Credits are correctly not charged on failure (confirmed all four times) — worth preserving as the fix ships.
3. Cross-tool contradictions on plain yes/no facts about the same site
CriticalOn the furniture site, AI Readiness Check (llms.txt: Missing), Full Site Audit (llms.txt: Missing — Opportunity), and llms.txt Generator (llms.txt: Present) gave three different answers to "does this site have an llms.txt file" — not a scoring-methodology difference, a factual disagreement, run back-to-back on the same domain. The Hermès test (finding #1 above) reproduces the same class of bug at larger scale and — critically — proves that when tools disagree, at least one of them is verifiably wrong, not just inconsistent. Reconciling this by averaging the tools or trusting whichever is loudest would not be a real fix.
4. AI Page Rewrite fabricates review-rating structured data
Critical · Compliance riskRunning AI Page Rewrite against a real product page generated ready-to-paste JSON-LD including specific aggregateRating values — "ratingValue": "4.5", "ratingCount": "1250" for the Organization, 4.6 / 245 for one product — despite the same audit's Landing Page Audit run on the identical page explicitly flagging "no customer reviews/ratings shown," and no tool in either test session finding any visible review system on the site.
If a user pastes this schema onto their live site as invited ("Copy / Download"), they'd be publishing fabricated review-count structured data — a violation of Google's structured-data guidelines for reviews, which can trigger a manual action or rich-result suspension sitewide. This needs an urgent check: does the tool ever fabricate aggregateRating or similar unverifiable fields when no real review data exists, instead of omitting the field?
5. Locale/currency defaults don't reflect the active project's actual market — three confirmed instances
Critical · New this sessionThis session found the clearest, most reproducible version of a pattern first spotted as "inconclusive" in the prior test round.
- Active project: Hermès USA (
hermes.com/us/en). Open Keyword Analysis. - Run "Keyword metrics" for "birkin bag" / "silk scarf," Location = United States.
Result: the CPC column returned S$0.38 and S$1.11 — Singapore-dollar values — despite Location being explicitly set to United States. Confirmed via screenshot this session.
Opened fresh on the same Hermès USA project. The Location field defaulted to Singapore, not United States — despite the active project's own domain being hermes.com/us/en and the project literally named "Hermes USA."
Combined with the third instance — Content Checker suggesting UK-pound rewrite copy for a Singapore business in the earlier test round — this is now confirmed on three separate tools, spanning both generated-content defaults and form-field defaults. It doesn't look like an isolated prompt issue; it looks like the active project's locale/market context isn't being threaded through consistently to every tool and every field that needs it. Worth an audit of exactly which fields and prompts receive project locale, and which fall back to a hardcoded default (Singapore appears to be that default, consistent with this account's original home market).
6. Output truncation — confirmed five times across four tools
Friction · SystemicBacklinks Explorer, SEM Ad Copy Generator, AI Content Optimiser, and AI Mentions Tracker all produced text that stops mid-sentence or mid-word — e.g. an SEM ad description reading "Free assembly on orders over." with the actual number missing, and a Backlinks Explorer recommendation cutting off inside the word "authority" on the Hermès re-test. Same symptom, four different tools, two different sites — almost certainly one shared output-length limit hit during generation. Fixing it once should fix all instances.
7. Competitor tools named by name in "Do this next" advice — confirmed 5×
Friction · TrustRank Checker, Page Technical & Domain Analysis, and SEO Strategy all told the user, unprompted, to go use Ahrefs, SEMrush, Moz, or Screaming Frog instead of Digimetrics' own equivalent tool (e.g. Rank Checker's own Keyword Analysis tool already does what it told the user to get from SEMrush). Almost certainly an unconstrained recommendation-generation prompt shared across tools — one guardrail fixes all five instances.
8. Raw backend variable names leak into user-facing copy — confirmed 4×
FrictionTechnical SEO Crawler (avgOnPageScore) and Backlinks Explorer (spamScore, referringIps, domainRank) both surfaced camelCase field names directly in generated sentences — e.g. "Your spamScore of 25 is moderate." Same fix pattern as #7: one shared prompt template, likely fixed once.
9. Stale form state carries over across project switches — confirmed twice
Friction · Agency-use riskSwitching the active project to Hermès USA and opening Competitors Identifier pre-populated the keyword field with sofa singapore — left over from the furniture project. This session reproduced the same class of bug on Keyword Analysis: switching to Hermès USA and opening the tool showed a stale media one keyword chip, even though the Website field correctly auto-filled to hermes.com. For a single-business user this is a minor annoyance; for an agency managing multiple clients — Digimetrics' core stated use case — running a tool without re-checking every field risks silently applying Client A's inputs to a report delivered under Client B's name.
03 Fix next
Real issues, narrower in scope than the cross-cutting findings above.
- DataHome dashboard's Site Health, Page Speed, Authority, and Technical Signals cards pull from a different project (
arcade.software) than the active one — confirmed a same-account scoping bug, not a cross-customer leak. - Silent failKeyword Analysis silently fails on large pasted keyword lists (~100+) with a generic error and no upfront warning about a likely batch-size limit.
- DataSEO Diagnostics' Keyword Difficulty pegs to exactly 100 for the large majority of keywords — including ones already ranking #1–4 — on both test sites. Reproduced at larger, cleaner scale on Hermès (100 keywords, same pattern: long-tail/local queries get real varied scores, head terms disproportionately hit the exact ceiling). Looks like a silent failure defaulting to max-difficulty, not genuine uniform difficulty.
- ContentMedia Plan Generator's "Marketing funnel" section is generic eco-friendly/sustainability boilerplate, completely disconnected from the actual business brief — reads like a hardcoded template not wired into the tool's context.
- ContentPersona Generator's summary undercounts its own output — reproduced on both sites. Furniture: "three distinct customer types" when 10 were generated. Hermès: "three distinct high-value customer types" when 11 were generated. The summarization step appears to see only a truncated slice of the full result.
- DataBacklinks Explorer's summary-vs-table numbers disagree on both sites (furniture: 63 vs. 57; Hermès: 426,256 vs. 333,570 — a ~28% gap). On Hermès the recommendation text also does its own arithmetic and gets it wrong: states "46%" for a ratio that computes to ~36% — a genuine calculation error, not just a data-source mismatch.
- LabelingTool count shown as "36" (sidebar, Home) vs. "35" (Tools page's own header and a manual tile count). Pick one source of truth.
- LabelingThree tools (
/seo-diagnostics,/performance-audit,/social-audit) use non-standard URLs and are missing their per-tool browser-tab title, unlike every/tool/*page. - LabelingSidebar nav item "Settings" actually links to the Account & billing page — no separate Settings page exists.
- DataAccount page's "X monthly" plan-allowance caption appears to track the live credit balance instead of staying fixed at the plan's real 6,000/month quota.
- UXSocial Media Audit rendered a stale "Brightpath" demo result as if it were a live run, before any input was entered — vanished on hard reload, so very likely stale client cache rather than a data leak, but still confusing as experienced.
- CopyFull Site Audit's recommendations are written as "Digimetrics will do this for you" rather than advice to the user — four instances read like an unedited agency-deliverable template, misrepresenting what a self-serve customer gets.
04 Worth doing, lower urgency
- PolishFive overlapping onboarding/gamification widgets stacked on Home (sticky banner, plan-step card, floating "Finish profile" pill, "Explore the platform" checklist, "Your plan" 6-step tracker) — consolidate to one primary mechanism.
- PolishPage Speed Check shows two "Largest Contentful Paint" numbers (field: 2.6s vs. lab: 21.6s) under the identical label with no explanation of why they differ so much.
- ConsistencyDepth of the "Recommendations & next steps" section varies a lot tool-to-tool — some get full priority/assignee-routed action plans, some get a plain numbered list. Decide if that's intentional.
- DataAnchor Text Cleaner's "Anchors to fix" table has several exact-URL repeats with no "found on page" column to confirm whether they're legitimate duplicates or a rendering issue.
05 What's already working
Worth protecting deliberately as the product scales, not just leaving alone.
06 Pricing findings
Full competitor pricing table, per-tool credit ledger, and market-rate recommendations live in the companion pricing report. Headline findings below.
The core story is bundling, not per-lookup price. A customer replicating Digimetrics' spread — SEO data + AI-visibility tracking + AI content drafting + ad-copy generation — across best-of-breed standalone tools would pay $350–1,800+/month (Ahrefs/Semrush $130–500 + Profound/Peec $99–399 + Surfer/Jasper $99–999). Digimetrics' Expert plan bundles equivalents of all four for $199/month. That gap doesn't come through clearly on the current pricing page, which emphasizes credits and tool count instead.
Per-run economics (pulled from the Projects & Runs ledger, the authoritative source): 1-credit lookups (Keyword Analysis, Rank Checker, Schema Generator, etc.) cost the customer roughly 3–5¢; 5-credit single-page analyses roughly 17–25¢; the two heaviest tools, Full Site Audit (50cr, ~$1.65–2.50) and AI Content Optimiser (12cr, ~40–60¢), are correctly the most expensive runs in the catalogue.
See the pricing report for the full per-tool recommendation table and the customer-price-vs-COGS distinction.
07 Full test log
Every tool run this audit, condensed. Tools with a real finding are detailed above; everything else is logged here for completeness rather than repeated as prose.
| Tool | Tested on | Credits | Result |
|---|---|---|---|
| Keyword Analysis | Both sites | 1 | Strong output; stale cross-project state + currency/locale bug — see #5, #9 |
| Rank Checker | Both sites | 1 | Ranks correctly (#1 for "birkin bag" on Hermès); named competitor tools by name — see #7; locale default bug — see #5 |
| Time to Rank | Furniture | 1 | Clean, works well |
| Anchor Text Cleaner | Furniture | 1–2 (est.) | Found a real issue (75 broken anchors); minor table de-dup question — §04 |
| Technical SEO Crawler | Both sites | 2 | Good streaming UX; variable-name leak — see #8; H1 claim ground-truth verified true on Hermès |
| Page Speed Check | Furniture | 1 (est.) | Runs correctly; confusing dual-metric labeling — §04 |
| On-Page Optimisation | Furniture | 5 | Best-in-class depth (108 headings, 84 images assessed); clean |
| Page Technical & Domain Analysis | Furniture | 5 | Handles missing data honestly; named competitor tools — see #7 |
| SEO Strategy | Furniture | 5 | Genuinely strategic output; named a competitor tool — see #7 |
| SEO Diagnostics | Both sites | 5 (est.) | KD=100 fallback pattern — §03 |
| Full Site Audit (SEO+AI) | Both sites | 50 | Deepest tool in the catalogue; cross-tool contradiction — see #3, #1; agency-voiced copy — §03 |
| Competitors Identifier | Both sites | 1 | Best-in-class; handled a nonsensical query gracefully — §05 |
| Backlinks Explorer | Both sites | 5 | Excellent real depth (921,847 backlinks on Hermès); truncation, variable leak, arithmetic error — see #6, #8, §03 |
| Schema Generator | Furniture | 1 | Instant; odd default image URL — minor |
| Social Media Audit | Furniture | 5 | Stale demo data rendered pre-input — §03 |
| Media Plan Generator | Furniture | 5 | Excellent personas/budget; generic funnel section — §03 |
| Persona Generator | Both sites | 5 | Excellent, market-adapted personas; summary undercount — §03 |
| Landing Page Audit | Furniture | 5 | Clean; discloses its own model (Claude Haiku) |
| SEM Ad Copy Generator | Furniture | 5 | Strong breadth; truncated ad copy — see #6 |
| Performance Marketing Audit | Furniture | 5 | Clean, useful context for the agency-voice finding — §03 |
| AI Content Optimiser | Furniture | 12 | Rewrite didn't complete; score formatting bug; truncation — see #6 |
| AI Page Rewrite | Furniture | 5 | Thorough; fabricated review schema — see #4 |
| AI Mentions Tracker | Both sites | 10 | Genuinely differentiated live 5-LLM tracking — §05; truncation — see #6 |
| AI Image Generator | Both sites | 4 (0 on fail) | Broken — see #2 |
| AI Video Generator | Both sites | 20 (0 on fail) | Broken, same root cause — see #2 |
| Content Checker | Both sites | 5 | Suggested £ for a Singapore business — see #5; Hermès re-test inconclusive (different test conditions) |
| Caption Generator | Furniture | 1 | Clean, no issues |
| Content Pillar Framework | Furniture | 1 | Clean, no issues |
| llms.txt Generator | Both sites | 1 | Correctly detects robots.txt on Hermès; part of contradiction — see #3 |
| AI Readiness Check | Both sites | 10 | Part of contradiction — see #3, #1 |
| Search Console / GA4 / Google Ads / Meta Ads / LinkedIn Ads | n/a | Free | Clean, honest empty-state handling; Meta/LinkedIn correctly marked "Coming soon" |
| Home, Account/Billing, Admin, Connect Data | n/a | — | See findings #9, §03; Admin is well-built internal tooling, customer list not reproduced here for privacy |
08 Engineering handoff
A companion file built for your engineering team (or an AI coding agent) to act on directly.
📄 digimetrics-fix-handoff.md
Every finding above, restructured as an ID'd, actionable engineering brief: exact reproduction steps, expected vs. actual behavior, and a root-cause hypothesis for where to look in code — for all 19 issues (9 critical + 10 fix-next/lower-urgency). It explicitly instructs whoever picks it up to verify each item is real before fixing it, not act on this document blindly.
- Reproduce the issue using the steps given
- Confirm the root-cause hypothesis against actual source (I never saw this codebase — my hypotheses are behavioral, not code references)
- Fix only what's confirmed, and check whether one fix covers multiple listed instances (several items explicitly share a root cause)
09 Methodology
Two test rounds. Round 1 used Furnituresg.com.sg, a small, thinly-indexed Singapore SME — a good test of sparse-data handling, but not of scale or technical sophistication. Round 2 (this update) used Hermès USA, a large, professionally engineered global luxury site with ~920,000 backlinks and likely meaningful client-side rendering, specifically to stress-test the tools against the opposite end of the spectrum.
Every factual claim a tool made about the live site (robots.txt, structured data, H1 tags, meta description, semantic HTML) was independently checked against the real page via direct DOM inspection in a real browser — not taken on faith. That verification step is what turns disagreements between tools into a provable claim about which one is actually wrong.
This version of the report replaces embedded screenshots with precise reproduction steps and verbatim quoted output, since the environment used to build this document doesn't support saving captured screenshots to an embeddable file. Every finding above can be independently re-run using the steps given.