SEO in YuktiBOS: audit, change list, approval, re-check (and the public checkup we haven't built)

The problem behind it

IIPTA, Mukika and the YuktiBOS site all need SEO, and SEO results take months. On 3 November I said start now. But I did not want to keep asking Claude to audit sites by hand. IIPTA's team has to run SEO themselves, with the same pattern we use for content: data, a page audit, exact changes, a human review, and then the change applied.

We already had a page review tool built on the content engine. It reviewed pages from a list of URLs. It was a base, but it had no crawler, no speed data, no exact fix list and no way to apply anything.

How it works today

From the help article:

1. A person presses Run audit. Reading and speed checks are free. Only AI wording costs money, and the real cost is shown beside the button.
2. The audit runs in the background: robots.txt and sitemaps first, then pages one a second (500 at most), then PageSpeed for the home page and the five most-linked pages, then the report. The bell rings when it is ready.
3. You read a score out of 100 and "Start here".
4. The change list: each line is one page, one element, what it says now, what to change it to, and why. You type the new text or press "Write wording with AI" (25 pages at most per run, estimate first). Approve or set aside.
5. Technical lines (robots.txt, redirects, canonical, noindex) are approved by the super admin only.
6. Create tasks (one Work task per page) or Do it for me (one task for the Mabbit team).
7. Re-check changes. Each line becomes "Fixed and checked" or "Still not fixed" with the reason.

There is also a panel for AI-search readiness (are AI crawlers allowed, do pages hold their text without scripts, is there structured data) and a panel that reads Search Console numbers already connected. The panel says plainly that Google states ordinary SEO is what matters for its AI features.

The decisions

Code finds facts, AI only words them, a human approves. The checks are written in code with fixed plain-language wording. AI never decides what is wrong. It only words titles and descriptions when asked, and code checks that text.

We never edit the client's site. Changes become Work tasks. The longer-term route is a pull request or a CMS write that the owner keeps, not a script that disappears when you stop paying.

A crawler that behaves. It identifies itself as YuktiBOS-SEO/1.0, obeys robots.txt, waits at least a second between requests, reads only same-site pages and checks every address, including each redirect hop, so it can never be pointed at a private network. The queue lives in the database, so a stopped run resumes within 90 seconds.

Reuse before building. The spec critique found that Search Console data and rank tracking already existed in the code. The spec had planned to build them again. It also found a new table that contradicted its own line saying to extend an existing one, and an unbounded AI wording cost: a 500-page audit could cost about Rs 2,500 for one site. The fix limited wording to pages with an issue and high inlinks, with the estimate shown first.

robots.txt advice we checked. Do not block CSS or scripts that build the main content. It is fine to block a script not needed to see the content, such as a chat box or tracking. robots.txt does not hide a page from Google; use noindex or a password.

The research that set the direction

We read 29 SEO products in six groups (separate post). The gap we chose: the full loop with a human approval, in plain language, for Indian small businesses and education businesses. The same research proposed a free public tool, a "YuktiBOS SEO Checkup": enter a URL, crawl up to 25 pages, get a score and the top five problems, with the rest behind an email.

What's not built

The public SEO Checkup page is on the next list. Before it launches it needs rate limits, a bot check, a daily spend cap, and a review under India's data protection law, which the research said it had not verified. I have not run the audit on iipta.com live. The paid tier price in the research (Rs 999 to 2,999 a month) is an assumption to test.

What went wrong

The page review tool had two live bugs when we looked: a settings page filtered for one job type while the site audit created another, and errors in the background job were only logged to the server, so a person saw nothing. Fixed first, before building on top.

Numbers at a glance