How to Review Hermes Agent SEO Output: Catch AI Slop, Verify Claims, Log Results
An AI agent with real SEO data still writes confident mistakes. This guide gives you a review routine for Hermes Agent output: confirm which tools actually ran, trace every number to its source, scan drafts for AI slop, check batches for scaled content risk, and keep a results log that feeds corrections back into your skills.
What counts as AI slop in SEO content?
AI slop is text that is fluent, on-topic, and empty: it could be pasted onto a competitor's site without changing a word, and it contains no fact a reader could not have guessed. In SEO it shows up as generic intros, unsourced statistics, keyword-stuffed headings, and advice that fits every website. Google's guidance does not treat AI-written content as spam in itself. It says using generative AI tools "to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse" (Google Search Central, "Using generative AI content", updated December 2025). The problem is missing value, not the tool.
How do you review SEO output from Hermes Agent?
Check which tools ran before reading the answer
Open the session and confirm the agent called web_extract, the browser, or an MCP tool before it made claims about a page or its data. The final answer will not tell you this; the tool calls will. Use hermes sessions browse to reopen a past run, including cron runs. A page audit with no fetch call is the model describing a page from memory, and a traffic figure with no Search Console call was invented. This check throws out the worst output before you spend time on it.
hermes sessions list hermes sessions browse
Label every claim by where it came from
Tag each factual statement as one of four types: measured (a figure from a tool call, with its date range), cited (a fact with a source URL), inferred (the agent's reasoning from the data), or unsupported (no source). Unsupported claims get verified or deleted, never passed along. Inferred claims are fine as long as they read as opinion, not measurement. You can have the agent do a first pass with the prompt below, but check its labels yourself; a model grading its own work tends to be generous.
Go through your previous answer. For every factual claim, add one label: [MEASURED: tool name + date range] [CITED: source URL] [INFERRED: which data it is based on] [UNSUPPORTED] Do not change or remove any claim. At the end, list every UNSUPPORTED claim.
Spot-check numbers against the source
Pick at least three figures and find each one yourself in the Search Console interface or the DataForSEO response, with the same date range, page, and filters. A mismatch is not always a hallucination, so look for scope errors first: a different date range, a property with and without www, a query filter the agent applied without saying so, or fresh data that has since changed (mcp-gsc includes non-final data when GSC_DATA_STATE is all). If one of the three is wrong and you cannot explain why, check every number in the output.
Open every URL the agent cites
Click each link and confirm it loads and says what the agent claims it says. Agents cite URLs that look right but do not exist, and real pages that do not contain the quoted statement. In SEO that costs you twice: a recommendation built on a misread competitor page is wrong, and a fabricated source that slips into published content damages the credibility you are trying to build.
Scan drafts for AI slop
Read any draft aloud and mark the patterns from the checklist below: filler openers, stock AI vocabulary, identical sentence rhythm, hype with no number behind it, and advice that fits any website. Then ask one question of every paragraph: could a competitor publish it unchanged? If yes, it needs a fact, example, or number only your site has, or it goes. Do not use an AI detector as the pass/fail gate. The question is whether the text is useful, and a detector does not measure that.
Confirm each recommendation on the live page
Before implementing a fix, open the page and confirm the problem exists. Agents recommend adding FAQ schema to pages that already have it, flag missing meta descriptions that are actually present, or suggest rewriting titles that changed last week. Implement a recommendation only when its evidence points at something you can see on the page right now.
Check batches for scaled content risk
If the agent produced more than a handful of pages, put them side by side and look at what actually changes between them. If only the city, product, or keyword changes and the rest is the same structure with swapped synonyms, you are close to Google's description of scaled content abuse. Each page needs something the others lack: its own data, examples, prices, or answers. Publishing fewer pages that each earn their place is the safer call.
Log the result and feed it back into the skill
Write a short results log entry for every run you act on: what the agent recommended, what you verified, what was wrong, what you changed, and when you will check the outcome. Then add each recurring error to the Pitfalls section of the skill that produced it, or ask Hermes to patch the skill with what it got wrong (template below). The log turns one-off corrections into rules, and the next run starts with them.
Measure the outcome with equal before-and-after windows
Once a change has been live for several weeks, compare equal date ranges before and after for the changed page, using mcp-gsc's compare_search_periods tool or the Search Console interface. Record clicks, impressions, CTR, and average position next to the change in your log. Keep the conclusion modest: a single page can move because of seasonality, a competitor, or a Google update, so note what else happened in that window before you credit the fix.
The AI slop checklist: what to look for
Filler openers
Delete on sight: "In today's fast-paced digital world", "When it comes to", "It's important to note that", "Let's dive in". The Indonesian versions are just as common: "Di era digital ini", "Tak dapat dipungkiri", "Mari kita bahas".
Stock AI vocabulary
delve, leverage, utilize, seamless, robust, elevate, unlock, harness, landscape, game-changer, cutting-edge. One is a coincidence. Four on the same page usually means nobody edited it.
Numbers with no source
"Studies show 70%", "up to 3x more traffic", "experts agree". If the draft cannot name the publisher and year, the number is either invented or unverifiable, and either way it comes out.
Advice that fits any site
"Create high-quality content", "optimize your meta tags", "focus on user experience". True everywhere, useful nowhere. Usable output names the page, the element, and the exact change.
Identical rhythm
Every paragraph exactly three sentences, every sentence about the same length, every section closing with a one-line recap. Human writing varies. Read it aloud and the pattern is hard to miss.
The "not just X, it's Y" habit
"It's not just a tool, it's a strategy." Once in an article it can land. As a recurring structure, it is one of the most recognizable AI tells.
Keyword-stuffed headings
The same exact-match phrase in the title, every H2, and the first sentence of every section. Put the keyword where it belongs (title, URL, first 100 words, one heading) and write naturally everywhere else.
Conclusions that repeat the intro
A closing section that restates the opening and adds nothing. If you can delete the conclusion without losing information, delete it.
What does usable agent output look like next to slop?
| Check | AI slop | Usable output (example) | Your action |
|---|---|---|---|
| Audit finding | Your meta descriptions could be improved | Meta description on /pricing is 212 characters and repeats the H1 word for word | Confirm on the live page, then fix |
| Traffic claim | This change could increase traffic by 30% | Clicks fell from 1,240 to 890 (1–7 Sep vs 25–31 Aug, Search Console) | Match the figure in Search Console with the same filters |
| Source | According to recent studies | Names the publisher and year, with a URL that loads | Click it and find the quoted claim |
| Recommendation | Add FAQ schema (it already exists) | No FAQPage JSON-LD found in the fetched HTML | Check the rendered page source |
| Content draft | Generic intro, stock phrases, advice for any site | Opens with the answer, uses facts only your site has | Cut paragraphs a competitor could copy |
| Batch of pages | 50 city pages where only the city name changes | Fewer pages, each with its own data or examples | Merge, or add unique content before publishing |
Templates for your review notes
## 2026-09-15 · /pricing · page-audit skill Agent recommended: - Shorten meta description (212 chars) - Add top query "harga paket bisnis" to the H1 Verified: - [x] Meta description length (checked on live page) - [ ] Query volume claim: no tool call behind it, removed Wrong / slop found: - Flagged a missing canonical; it is present in the server-rendered HTML Changed: - Meta description rewritten, H1 updated, live 2026-09-16 Before (Search Console, 2026-08-18 to 2026-09-14): - clicks 890 · impressions 24,100 · CTR 3.7% · avg position 8.2 Recheck: 2026-10-14 (same 28-day length after the change) Skill update: - Pitfall added: check canonical in fetched HTML before flagging it missing
The figures are placeholders showing the format, not real data. The line that matters most is the last one: every wrong finding should become a pitfall in the skill.
Update the page-audit skill. Add these to its Pitfalls section, one line each, written as instructions: 1. You flagged a missing canonical on /pricing, but it is present in the server-rendered HTML. Check the fetched HTML before flagging a canonical as missing. 2. You stated a search volume figure without a DataForSEO call. Never state search volume without a tool result. Show me the change before saving.
With skills.write_approval set to true, Hermes stages skill edits for your review instead of saving them straight away.
You are reviewing SEO content written by another AI. Be strict. For the draft below: 1. Quote every sentence that could appear unchanged on a competitor's site. 2. Quote every number and say whether a source is given. 3. Quote every phrase from this list: delve, leverage, seamless, robust, unlock, game-changer, in today's digital world, it's important to note, di era digital ini, tak dapat dipungkiri. 4. List claims a reader cannot verify. Do not rewrite anything. Only report. [paste draft here]
Run it in a new session, or with a different model, so the reviewer has not already seen and accepted the reasoning behind the draft.
What this review routine cannot catch
A clean checklist does not prove content is accurate or that it will rank. The slop scan checks how text reads, not whether its facts are true; only steps 3 and 4 test facts, and only for the claims you check. Spot-checking three numbers finds systematic errors, not every error. A before-and-after comparison on one page cannot separate your change from seasonality or a Google update. And no routine replaces knowing the subject: if nobody on your team can tell whether a recommendation is right for your market, the output should not ship.
Frequently asked questions
See how AI answers describe your brand today
Intura sells AI Search Optimization Services, so read this with that in mind. Hermes Agent helps with the pages on your own site. To see whether ChatGPT, Gemini, Claude, Perplexity, and Google AI Overview mention your brand at all, our free AI visibility audit returns a score and a prioritized fix list. New to this series? Start with guide one on installing Hermes Agent.
Run a free AI visibility auditKey takeaways
Review Hermes Agent SEO output in the order that throws out bad work fastest: confirm the tools ran, label every claim by source, spot-check numbers against Search Console with identical filters, and click every cited URL. Then scan drafts for slop and batches for scaled content risk. Implement only what you can see on the live page. Log each run, measure with equal before-and-after windows, and write every recurring mistake into the skill that made it, so the same error does not come back next week.