We Audited 174 AI Character Pages

A public AI character page should not become indexable just because its database record says public and approved. We learned that by auditing all 174 character pages that our old search rule would have submitted to Google.

The rule looked reasonable on paper. A page had to be public, approved, suitable for the public surface, and outside one excluded import source. Official characters could pass immediately. Other characters needed at least one chat.

The data did not support that shortcut.

What the audit found

We reviewed the exact 174-page set on August 30, 2026. The audit checked the character name, introduction, persona, opening message, categories, images, source metadata, engagement counts, and the text rendered on the public page.

Here is the result.

• 169 pages came from an external import source.
• 70 still contained raw {{char}} or {{user}} template tokens.
• 32 exposed those tokens in the opening message itself.
• 5 contained an outside-account promotion.
• 4 contained a raw external URL in public character copy.
• 170 had no completed chats on Soliloquy.
• 33 carried more than 4,000 characters of profile copy, and 7 exceeded 8,000.
• 16 had no category, and 1 had no usable image.

Every page had an opening message. That sounds encouraging until the opening is treated as content rather than a checkbox. An opening can exist and still begin with an unresolved variable, an imported promotion, a wall of platform-specific formatting, or a scene that gives the reader no clear move.

We also found repeated description copy across multiple pages. Exact duplication was not the largest problem in this batch, but it was another warning that record-level approval and page-level editorial quality were doing different jobs.

Why public approval was not enough

Moderation approval answers a safety and publishing question. It does not certify that a page is useful in search.

An imported character may be allowed on the product while still being a poor search result. The page can contain a source platform's call to action, template syntax that our renderer does not replace, or a long internal persona written for a model rather than a human reader. None of those problems necessarily make the character unusable inside the app. They do make the public page look unfinished.

The word official created another false sense of confidence. In this dataset, official mostly described how a seed record entered the catalog. It did not mean an editor had read the final Soliloquy page.

Chat count was not a reliable substitute either. Only four of the 174 pages had any chats, and several of those four still showed an outside-platform reference, promotional copy, or raw formatting. Engagement can help prioritize a review. It cannot replace one.

The decision we made

We submitted the main sitemap, which currently contains 12 reviewed site pages. We did not submit the character sitemap.

We also stopped treating official status or a minimum chat count as automatic permission to submit a page for indexing. The next implementation step is a separate manual search flag that defaults to off. Until that control is deployed, we are not submitting the character sitemap or requesting indexing for individual character pages.

That distinction matters. Removing a page from a sitemap is not enough when it is already linked elsewhere. The page itself needs a clear indexing decision, and category pages with no approved characters should not be promoted as useful search destinations.

No character from this first batch was approved for submission. That is deliberately conservative. The audit found several plausible candidates, but none had enough evidence to justify skipping a complete editorial read.

The review standard from here

A character page can be considered for search only when a reviewer can answer yes to all of these questions.

1. Does the public title identify a specific character or scenario without imported platform labels?
2. Does the reader-facing introduction explain the premise in plain language?
3. Are all template variables resolved or intentionally supported by the renderer?
4. Is the opening message free of promotions, account handles, and unrelated URLs?
5. Does the opening establish a scene and leave the user a meaningful next move?
6. Is the visible copy formatted for a person, rather than pasted from a model configuration file?
7. Are the image, categories, canonical URL, and page description complete?
8. Has a human read the rendered page, not just the database fields?

The last question is the important one. A clean source record can still render badly, and a messy source record can be cleaned into a useful page. Search engines and readers see the result.

How we tested

We queried the production public dataset using the same eligibility rules as the character sitemap. We then scanned the five public text fields for unresolved character and user tokens, URLs, promotion phrases, residual HTML, and length outliers. We separately checked source distribution, image and category completeness, chat counts, and duplicate description groups.

Automated checks produced the counts above. We then opened a sample of flagged and apparently clean pages in the live site and compared the rendered copy with the underlying fields. This second step caught the core mistake in the old rule. A high content score or an official source label did not prove that the page was ready for a reader.

This is a baseline, not a permanent verdict on the characters. The point is to make approval explicit, repair the best pages one by one, and measure what happens after genuinely useful pages enter search.

The next two guides explain the editorial side of that work, starting with how to write a stronger AI character profile and how to write an opening message that gives the user something to do.

You can also browse the current character catalog. Public availability and search approval remain separate decisions in our review process.