SEO Agent: Same Eight-Module Rules as Web and CLI

By William Zhu · Cofounder, InfiniSynapse · Last updated: 2026-08-19 · Last verified: 2026-08-19 · Methods: eight-module page audits plus caller-versus-paste parity on live URLs — not a claimed Google score.

Author / off-site profiles: GitHub @allwefantasy · auto-coder · GitHub @InfiniSynapse · LinkedIn company · Editorial standards. Formal public work: InfiniSQL, auto-coder, retrieval systems. Desk: shipping SEO Health and the /en/tool/ visibility pages. No personal LinkedIn, award, or vendor badge.

Trust / COI: About · Corrections · Publishing principles · Privacy · NIST Privacy Framework · Vision. This site does not publish a standalone /en/terms URL; the editorial-standards page is the policy home. InfiniSynapse ships SEO Health as a credit-based desk; first-party counts are labeled; product CTAs are commercial.

Fact-check: NIST — AI · OECD AI Principles · CISA — AI · W3C Web Share · Stanford HAI AI Index 2025 · G2 SEO tools · Gartner Peer Insights · AgentSpot listing. Corrections: zhuhl@infinisynapse.com.

Dates (match schema): First published 2026-08-16. Last modified 2026-08-19. Desk run 2026-08-11. Last verified 2026-08-19.

TL;DR

Direct answer: An seo agent is a caller that fetches one live URL and applies the same eight-module rules as the web paste and the CLI — title, meta, headings, density, images, links, tech, and speed — then returns traffic lights and a next edit, not a second scoring philosophy.

What you'll learn

• A 46-word definition of an seo agent you can quote in a ticket
• Why the caller must reuse the web and CLI rule list
• What to pass in and what must come back
• How lights become an assigned edit
• Failure modes we refuse to call a pass on this desk

If the URL is already open, run a free page check before you let a caller invent a tenth module. The report only helps once the lights match the paste box.

What an agent is allowed to do

Key Definition: An seo agent is a caller that fetches one live URL and applies the same eight-module rules as the web paste and the CLI — title, meta, headings, density, images, links, tech, and speed — then returns traffic lights and a next edit instead of inventing a second scoring philosophy.

Figure. Public desk series DESK-AGENT-20260819A on 20 English /en/tool/ pages (run 2026-08-11, verified 2026-08-19). Caller red / 20: paste/caller parity 4, named next edit 7, eight modules named 5, request URL printed 3. Named next edit was the weakest caller check. Not a customer lift study.

Key terms

People type seo agent when they want a process that can check a page without a human pasting the address every time. That is a caller job. It is not a new exam. It is not a second product with a prettier score.

Public desk method: 20 pages, caller checks plus eight modules

First-hand, dated, reproducible — not a customer case study. We scored 20 English /en/tool/ addresses with the same eight-module rules, then scored how a caller would have to report those lights. Run date 2026-08-11. Last verified 2026-08-19. Next public re-run 2026-08-25. Marker DESK-AGENT-20260819A. Download desk-caller-n20.csv, desk-modules-n20.csv, and the pillar cluster file desk-cluster-n10.csv.

Judging rules. Green = the check has enough evidence to leave the file. Amber = borderline. Red = missing or mismatched. We do not invent a before/after lift from those bars. We do not print an official Google agent score.

On the same 20 pages, the eight modules the caller must reuse scored as follows. Links were the weakest page module.

The same morning we scored the 10 English on-page cluster URLs. Person author in JSON-LD: 0 / 10. Team-only hero byline: 10 / 10. Illustrative chart: 10 / 10. Dated first-hand block: 0 / 10. This URL (seo-agent) was Organization-authored and team-bylined. That is why a page about an seo agent starts with a Person and a downloadable series.

First-hand review: this URL on 2026-08-19

On 2026-08-19 I re-opened this live page and two sibling products: SEO CLI and on-page SEO checker. The 2026-08-11 cluster file still lists this slug as team-bylined with an illustrative chart. That row is historical; we do not rewrite it. Today’s body is a Person byline (William Zhu), three downloadable CSVs, and this dated paragraph. The model did not sit the 2026-08-11 scoring. I did. The surprise: named next edit was red on 7 / 20 caller rows while title was red on only 2 / 20 pages. A caller can reprint eight greens and still fail if it never names the edit. A stranger can reopen those files. This is a desk review of our own pages, not a third-party award. We do not publish a fake customer before/after from this desk.

Industry context used as evidence, not as a plaque: the Stanford HAI AI Index 2025 reports organizational AI use at 78% in 2024. Cheap drafts multiply. They do not create a named next edit. McKinsey’s State of AI still separates experimentation from production value. A caller that invents an “82” is still in the experiment column.

Independent reviews, directories, and specs

Third-party URLs a reviewer can open — retrieved 2026-08-19. None is an award, a vendor badge, or an seo agent grade.

AgentSpot is a directory mention of the company. G2 and Gartner Peer Insights are where independent reviews of adjacent categories live. NIST, OECD, and CISA are the governance specs a caller should map to. An seo agent becomes citable when those URLs stay dated and the CSV stays downloadable. Inventing a plaque would make the authority worse.

A useful seo agent answers three questions. Which URL did it fetch. Which of the eight modules are red. What is the next edit a person can ship in the same sitting. If the output is only “the page looks fine,” you still have the original problem.

The unit of work for an seo agent is one address. Not a domain average. Not a crawl of pages nobody will fix. You hand the caller a URL, it requests the live HTML, and it scores the same eight modules the paste box already uses. Software can run the fetch. A person still has to edit. Write the findings as module, light, and next edit. If you cannot name the edit, the seo agent is not finished.

The Wikipedia page on intelligent agents (Wikidata Q220659) is the right plain definition: a system that observes and then acts toward a goal. The goal here is a punch list, not a vibe. An seo agent that observes the HTML and then writes a new rubric has left that definition. IBM’s generative AI overview is a reminder that generation is cheap. Observation plus a named edit is the job of an seo agent.

Web paste, CLI, and a caller are three doors. They are not three tests. The category frame around those doors lives in On Page SEO Tool. The scriptable door is SEO CLI. The human paste door is on-page SEO checker. This page stays on the caller. Do not let an seo agent become a second exam because the invocation was a prompt.

The same eight-module rules

Walk the modules in the same order the paste box uses. An seo agent that opens on a prose summary and hides a noindex tag is entertaining you. On this desk, links were red on 9 / 20 pages while tech was red on only 2 / 20. A paragraph about “solid fundamentals” hid the dead hrefs.

Title, meta, headings, density

Title asks whether the document name can become a usable title link. Google documents how it may rewrite titles in its title-link documentation. Meta asks whether the description is a plausible snippet (snippet documentation). Headings should outline the page. Density is a band, not a quota. An seo agent that tells you to paste the same phrase into every H2 is stuffing wearing a caller.

If the caller cannot name which of those four modules is red, it is summarizing. Summaries are not lights. Demand the module name. On this desk, headings were red on 6 / 20 and density on 5 / 20.

Images, links, tech, speed

Images need real alt text (W3C WCAG 2.1). Links need live targets. Tech covers HTTPS, indexability, and a coherent head. Confirm fetch first with an SEO URL checker if eligibility is in doubt. Speed here is hygiene, not a rank forecast. An seo agent that skips dead hrefs because “the copy looks strong” has dropped half the rule list. On this desk, images were red on 7 / 20 and links on 9 / 20.

schema.org SoftwareApplication is evidence that a checker is a product with a name, a URL, and a declared job. Declare the job as eight modules. Do not declare a ninth “agent score.”

After the caller returns, paste the same URL into the human paste box and compare lights. If they disagree, believe the mismatch, not the prettier paragraph. On this desk, paste/caller parity was red on 4 / 20.

How the agent should call the check

A finished seo agent is a punch list for one address, not a chat that restates the H1. Sort by light. Fix reds before ambers. Leave greens alone.

What you pass in

Pass the exact URL a person would share: scheme, host, path, and query if the query is part of the document. Do not pass a homepage and hope the caller invents the article path. Do not let the caller collapse http and https into one row without showing the hop.

If you are choosing the scriptable door rather than the caller, that is an SEO CLI question. The rules should stay the same. Only who invokes them should change. Buyer surfaces around those doors sit in onpage SEO software.

Sharing the report is a separate step. The W3C Web Share API is how a browser hands a URL to another app. An seo agent should be that honest about the address it fetched. Print the request URL. Do not hide it inside a story. On this desk, request URL printed was red on 3 / 20.

What you must get back

You must get eight module names, a light per module, and one next edit you can assign. JSON is fine. Prose is fine if it still names the module. An seo agent that returns only “improve E-E-A-T” on an on-page pass has jumped pillars. On this desk, eight modules named was red on 5 / 20 and named next edit was red on 7 / 20. Those two rows are why an seo agent cannot stop at a paragraph.

We do not claim a Google on-page API. We report what the document declared against a written list. If Search Console later disagrees about clicks, believe Search Console for traffic and believe an seo agent for “what did we ship.”

Governance is evidence, not decoration. NIST’s artificial intelligence pages require you to map the risk, measure the leftover, and name an owner. That is why an seo agent that cannot name who edits the red has already failed the NIST posture, even if the prose is cheerful.

What the agent does with lights

Treat any single number as a qualitative estimate. Averaging this desk into one “74” would hide the 9 red link rows and the 7 missing next-edit rows. An seo agent that files a PDF and never changes the title is theatre.

OECD AI Principles put accountability on the organization, not on the model. Use that as a disclosure posture: say what it fetched, what it scored, and who edits. Hidden generation of a new rubric is a trust leak on the report itself.

Where this sits next to CLI and paste

The eight modules are the core. Two adjacent doors sit next to the caller. They are not substitutes.

CLI is the script; the agent is the caller

A CLI is the command a pipeline can hang on a pull request. An seo agent is the process that decides to run that command, read the lights, and open a ticket. If you only needed a script, stop on SEO CLI. If you needed a caller that will not invent modules, stay here.

Do not split the rule list because the invocation was a prompt instead of a flag. Prompt or flag, the exam is the same.

Paste is still the human surface

A person still needs a paste box they can trust. That transactional check is on-page SEO checker. Use it to verify the caller. If the human paste and the caller disagree, the caller is wrong until you can explain the delta.

The most common miss on this desk is not a missing title. It is an seo agent that restates the H1, adds a cheerful number, and never names a module. CISA’s AI page is the evidence we use for that row: automated systems still need an owner. If nobody can say who edits the red, the caller produced a memo.

A perfect JSON report that nobody opens is the second miss. The seo agent did the fetch. The team did not. If there is no rewrite in the same sitting, you ran a report, not a check. A third miss is letting the caller skip eligibility. Title polish on a noindex URL is costume. Tech is still module seven.

How to run one caller pass

Do not wait for a crawl film. An seo agent is a one-URL caller. The loop is the same six steps in the HowTo on this page. Keep the seo agent on that loop; do not expand it into a second scoring philosophy.

1. Name the exact URL. Confirm scheme and host. Print the request URL.

2. Invoke the same eight-module check the paste box uses. Do not invent a tenth module.

3. Read lights in order. Stop on tech reds before you rewrite density. On this desk, tech was red on 2 / 20; links were red on 9 / 20.

4. Assign every red to a person. Write the next edit in one sentence. On this desk, named next edit was red on 7 / 20.

5. Edit. Re-invoke. Compare the same eight lights. If they disagree with the paste box, the seo agent is wrong.

6. Only then schedule the caller on the next URL that shares the template. A new URL cannot prove the edit landed.

If step 5 does not change the HTML, you ran a report, not a check. Keep the punch list next to the tab. Close the tab only when the reds are gone or dated. A useful seo agent prints the request URL on every run so you can prove which address was fetched.

Failure modes

These are page and caller failures, not runtime-reliability sermons.

1. A new score per door. The paste box shows three reds. The seo agent writes “looks healthy.” You do not have parity. You have two products. On this desk, parity was red on 4 / 20.
2. Prose instead of a module. The caller never says title, meta, headings, density, images, links, tech, or speed. Demand the names. An seo agent that never says those eight words is summarizing.
3. Edit never happens. The JSON is pretty. The template is untouched. The seo agent did its job. The team did not.

None of these are “the model was unreliable.” They are rule-list failures. Fix the URL, or fix the caller.

Frequently Asked Questions

Is this a different product from the web paste?

Bottom line: No. It is a caller of the same eight-module rules.

• If the lights disagree with the paste box, the caller is wrong until you can explain the delta.
• On this desk, paste/caller parity was red on 4 / 20.
• An seo agent that invents a second exam has already failed this question.

Do I need a CLI first?

Bottom line: You need a rule list first.

• A CLI is the scriptable door.
• An seo agent is the process that invokes that door.
• If you cannot name the command, you are not ready to automate the paste.
• The seo agent still has to return the same eight lights the CLI would print.

Can the caller add extra modules?

Bottom line: Not on this pass.

• Extra modules become a second exam.
• Keep title, meta, headings, density, images, links, tech, and speed.
• Adjacent jobs get their own URL.
• An seo agent that adds a ninth “agent score” has left this pass.

What should I do with a vanity 100?

Bottom line: Ignore it. Read the reds.

• Averaging this desk into one “74” would hide 9 red link rows.
• An seo agent that only ships a doughnut is optimizing for a screenshot.

When is the caller finished?

Bottom line: When a person can name the next edit and the re-run matches the paste box.

• A paragraph that never names a module is not finished.
• On this desk, named next edit was red on 7 / 20.
• That sentence is the whole job of an seo agent.

Conclusion

An seo agent is a caller of the same eight-module rules the web paste and the CLI already use. Fetch one URL. Return lights. Name the next edit. Believe the reds. Ignore a decorative 100. Start with a free page check. For the tool frame around those modules, stay on On Page SEO Tool. The desk rows stay public so this seo agent can be cited; they are first-party counts, not a third-party award.

References

1. Google — Title links · Google — Snippets · W3C WCAG 2.1 · W3C Web Share API · schema.org SoftwareApplication. Retrieved 2026-08-19.
2. NIST — Artificial intelligence · NIST Privacy Framework · OECD AI Principles · CISA — AI.
3. Stanford HAI — AI Index 2025 (organizational AI use 78% in 2024) · McKinsey — The state of AI · IBM — Generative AI · Intelligent agent · Wikidata Q220659 · Search Engine Land (beat, not “as seen in”).
4. G2 — SEO tools · Gartner Peer Insights · AgentSpot — InfiniSynapse (directory mention, not an award).
5. InfiniSynapse desk — 20-page caller series, eight-module series, and 10-URL cluster (DESK-AGENT-20260819A); first-hand re-open of this URL on 2026-08-19. Not a customer before/after study.

About the author — William Zhu, cofounder of InfiniSynapse. Formal public work: InfiniSQL, auto-coder, retrieval systems (GitHub @allwefantasy). Reviewer: InfiniSynapse Data Team. Published 2026-08-16. Updated 2026-08-19. Policy: About · editorial standards · privacy. No standalone /en/terms URL.