---
title: "What An SEO Platform Cannot Catch: Building The Trust Layer Around Conductor"
url: https://www.axelerant.com/blog/what-an-seo-platform-cannot-catch
published: 2026-09-21T16:41:11.713Z
author: "Sayan Mallick" (sayan.mallick@axelerant.com)
source: Axelerant Thinking
---

# What An SEO Platform Cannot Catch: Building The Trust Layer Around Conductor

> Conductor scores content for search. It cannot catch affiliate links, missing medical disclaimers, or contextual errors. See the guardrail layer around it.

A global endurance sports brand was launching a new event series for women athletes. The supporting content came from guest contributors with doctorates: dense, well-cited, and full of sensitive medical territory.

Conductor gave every piece a content score. It could not see an affiliate link routing revenue elsewhere, a pregnancy article with no disclaimer, or the word “semester” where “trimester” belonged.

This is the story of the nine gaps we found, the guardrail layer we built around the platform to close them, and why the client’s original hope for a zero-touch pipeline was the wrong goal.

## The Problem: Optimized Is Not The Same As Safe

The brand was launching a new event series built for women athletes, and the content plan around it was ambitious: 29 articles on topics the site had never covered properly.

- Training through pregnancy.

- How the menstrual cycle affects race preparation.

- Female physiology and fueling.

- Returning to sport postpartum.

Most of the pieces were written by guest contributors with doctorates in exactly those fields. They arrived long, technical, heavily referenced, and, in places, wrong in ways that only a careful editor would notice.

The brand’s use of [Conductor](https://www.conductor.com/) had, by this point, shifted. The platform had been bought for content ideation and drafting. What the team actually needed from it on this series was editorial quality assurance: is this piece accurate, is it on-brand, is it safe, can it go live.

Conductor is very good at what it was built for. It covers keywords, content score, rankings, cannibalization, and briefs, and it does all of that well. It was not built to be a copy editor or a fact-checker, and the vendor confirmed as much when we raised it. That distinction matters when teams connect [discoverability work](/what-we-do/revops-and-growth/discoverability) to editorial governance.

The client’s expectation, understandably, was closer to zero-touch: hand the articles to the system, get back publishable pages. On a 16-week training plan, that expectation is almost reachable.

On an article about training in the second trimester, it is not. The cost of being wrong is not a lower content score. It is a reader acting on bad medical guidance under the brand’s name.

We did not have months. The series had a launch date, the client’s content lead was the sole approver and already stretched, and the platform’s writing assistant would not even start on several pieces.

It requires seven competitor URLs for analysis and, on niche topics like these, there were not seven credible competitors to find. The constraint was clear: we needed a review layer that caught what the platform could not, produced a short list of decisions for one busy human, and did it 29 times in a few weeks.

![Diagnostic comparing a Conductor score of 92 with three publication blockers: an affiliate link, a missing medical disclaimer, and a contextual terminology error](https://cdn.sanity.io/images/d78h1a2t/production/2f41945dd8491327ba461cf685cbf2637cfb7118-2000x1250.jpg)

*A high content score and three publish-blocking issues on the same article.*

## What We Did And Why

### First, Name The Gaps

We went through the first batch of contributor drafts by hand and cataloged every issue that Conductor’s scoring had not surfaced. Nine categories came out of that exercise, and they have held up across everything we have reviewed since.

- External link governance: affiliate tags routing revenue and attribution to other publishers, competitor and self-promotional links, tracking junk in URLs, dead links.

- Fact and claim verification: statistics, records, superlatives, and expert credentials stated as fact with no source.

- Sensitive and YMYL handling: missing support signposts on hard topics, unsourced medical claims, weak or absent disclaimers, stigmatizing framing.

- Consent and ethics: personal stories published without confirmed consent, pseudonyms handled inconsistently.

- Brand nomenclature and voice: wrong race names, wrong casing, drifting point of view.

- Internal logic: a “10 things” list with 13 items, self-contradiction between sections, two intents fused into one piece.

- Contextual knowledge: the word “semester” used for “trimester” in a pregnancy article, spelled correctly, grammatically fine, and completely wrong.

- Production artifacts: leftover editorial notes, copy-paste residue in bylines, encoding slips.

- Dense scientific content that needed to be preserved, not simplified away, while still being made safe for a general reader.

None of these are search problems. All of them are trust problems. A content score is indifferent to every one of them, and it should be. That is not the platform’s job.

### Then Build The Layer That Owns Them

We encoded the nine categories into our in-house agentic content system as a guardrail pass that runs on every draft before a human sees it. It sits between the contributor’s draft and Conductor’s scoring, a practical extension of the [AI-to-CMS last mile](/blog/ai-content-platform-cms-last-mile-drupal).

It is built on a principle we hold to strictly: flag, don’t fill. The system is allowed to identify a problem and propose a fix. It is not allowed to silently make a change on a claim it cannot verify, and it is never allowed to invent a source, a statistic, or a disclaimer’s medical content.

Everything it cannot verify becomes a flagged item with a severity and a recommended action, and the human decides.

In practice, the pass does several things in sequence:

- It audits every external link, classifying each as editorial, affiliate, competitor, or broken, and strips tracking parameters.

- It extracts every factual claim and checks whether a source is present, marking unsourced claims for the editor rather than removing them.

- It applies a house-style dictionary for race names, distances, casing, and terminology, plus a contextual check that reads terms against the topic. That is how “semester” got caught.

- It enforces a mandatory disclaimer rule for any piece touching training load, nutrition, or medical conditions, using approved disclaimer language the client supplied.

- It checks internal consistency: list counts, contradictions between sections, headings that promise one thing and deliver another.

- It runs a final pre-delivery checklist covering US spelling, no generic headings, and the removal of production artifacts.

The output for each article is a critique with severities and a verdict: publish, publish after these fixes, or hold for a human decision. For a series like this, most pieces landed in the middle category with two to five flagged items.

### Where Conductor Fits After That

Once a piece cleared the guardrail pass and the client’s content lead had resolved the flags, it went into Conductor for what Conductor does well: scoring against the brief, cannibalization checks against existing pages, and heading structure against what ranks.

Where the writing assistant could run, it ran. Where it stalled on the competitor-URL requirement, we skipped it and relied on the brief and scoring only. We fed every one of those stalls back to the vendor as product feedback, along with the request for multiple content profiles per draft, since several of these articles served more than one audience persona at once.

The alternative we rejected was to keep absorbing the gap quietly by having our own editor do all of this by hand. We had been doing exactly that on the earlier refresh work, and it was not visible to the client, not scalable, and not honest about where the value was actually being created.

Making the layer explicit and repeatable changed the conversation from “why is the platform not doing this” to “here is the division of labor that works.”

![Six-stage architecture from contributor draft through guardrail checks, flagged decisions, human approval, Conductor scoring, and publication](https://cdn.sanity.io/images/d78h1a2t/production/2f4878a0c26a1bcf615a15415b0350f6301cf13a-2000x1250.jpg)

*The guardrail layer separates trust decisions from search scoring.*

## What Changed

All 29 articles went through the guardrail pass for the launch, with the client’s content lead approving from a flag list rather than from a full read of every draft. That is the output that mattered most to her.

The review load per article dropped from reading two to three thousand words of dense science to resolving a handful of specific decisions, each with the passage quoted, the issue named, and a proposed fix attached.

Three catches from the first batch became the examples we now use to explain the layer to anyone evaluating the platform:

- An affiliate link a contributor had included, routing revenue and attribution to another publisher, which nothing in the scoring would ever flag.

- A pregnancy-and-training article with no medical disclaimer, on a topic where the brand’s own policy requires one.

- “Semester” for “trimester,” a correctly spelled word that no spellchecker or content score would question and that would have been embarrassing in print.

Each was found by a different check in the pass, which is why the pass has nine checks and not one.

The vendor relationship improved rather than soured. Every stall and gap went back to the product team as structured feedback: the seven-competitor requirement blocking niche topics, the single content profile per draft, and the existing URL showing up in its own competitor list.

More importantly, the client stopped measuring the platform against a job it was never designed for and started measuring it against the job it does.

### The honest edge.

This is not zero-touch, and it will not become zero-touch. The client wanted a system that took a contributor draft and returned a publishable page with no human in the loop. On sensitive content, we do not think that goal is responsible, and we said so.

What the layer does is make the human’s time count: one approver, a short list of real decisions, no rewriting. It also does not verify facts against the world. It verifies whether a source is present and whether the claim is internally consistent.

Checking that the source actually says what the article claims is still a human job, and on medical content it should be.

The layer also has to be taught. The nomenclature dictionary, the disclaimer rules, the persona voice, and the list of approved sources were built up over the first two months of the engagement. The pass got sharper as it went.

A team adopting this on day one should expect the first ten articles to surface gaps in the rules as often as gaps in the drafts.

![Content review artifact reducing a source draft to four evidence-backed decisions and a publish-after-fixes verdict](https://cdn.sanity.io/images/d78h1a2t/production/4c74670653465bc346bbe4434e29edae7f22595e-2000x1250.jpg)

*What the content lead reviews now: the flags, not the draft.*

## What You Can Take From This

Decide what your platform is for before you judge it. Conductor, and every tool in its category, is built to make content discoverable. Trust, accuracy, brand safety, and legal exposure are a different job with different failure modes, and a high content score tells you nothing about any of them. Teams that bought a platform expecting it to be an editor end up disappointed with a tool that is doing exactly what it was designed to do. The same separation underpins [content operations in a composable platform](/blog/composable-dxp-content-operations).

Write your nine gaps down. Ours are above and you are welcome to start from them, but the useful version is the one built from your own drafts. Take ten recent pieces, list every issue a careful editor would have caught that your tooling did not, and group them. That list is the specification for your guardrail layer. It can be built with an agentic system or with a checklist and a person. The exercise pairs naturally with a [data-led content refresh](/blog/reviving-dead-content-library) when teams are deciding what to keep and what to trust.

Then adopt flag, don’t fill as a hard rule. The temptation with any automated editorial system is to let it fix things. On sensitive content, that is how invented sources and hallucinated disclaimers get published under your name.

A system that flags with a proposed fix and stops is slower by a few minutes per article and safer by an order of magnitude. Your single approver will thank you, because their job becomes deciding rather than reading. That same discipline matters when content is being structured for [AI search citation](/blog/why-being-famous-does-not-get-you-cited-in-ai-search).

## See What Your Content Score Cannot

Publishing expert or contributor content on sensitive topics? [Send us three pieces](/contact) and we will come back with what a guardrail pass would have flagged.

---

Read on the web: https://www.axelerant.com/blog/what-an-seo-platform-cannot-catch