SEO and AI Search Improvement Practice Log

Tomohiro Iida · Published June 25, 2026 · Updated August 6, 2026

This page is a practice log recording the SEO and AI search changes we made on netsujo.jp, without narrowing it to the successes. For each measure we separate what we did, what we could observe, what cannot be claimed as cause, and our current assessment, and we update it over time. Where a past judgment turned out to be wrong, we do not simply delete the old wording — we keep the correction date and the reason for the change.

netsujo.jp is a young site and its traffic is still very small. There are no dramatic results, and we do not write that results exist when they do not. The value of this log is not a story about what worked; it is in laying out separately what we did, what we observed, and what can and cannot be stated as cause.

One premise for the whole log, stated first. In its official guide, Google says that optimizing for generative AI search is optimizing for the search experience — in other words, SEO itself. There is no special additional requirement for appearing in AI search; the condition is that a page is indexed and eligible for snippet display. On that premise, much of what we had treated as an "AI-only measure" needed its positioning reviewed.

Source: developers.google.com/search/docs/fundamentals/ai-optimization-guide

Key takeaways

This page is a practice log recording the SEO and AI search changes made on netsujo.jp, without narrowing it to the successes. Each measure is updated as what we did, what we could observe, our current assessment, and what cannot be claimed as cause.
Volume of work and number of pages published are productivity information; search traffic, mentions in AI, and inquiries are outcome information. We keep the two apart.
Where a past judgment was wrong, we do not delete the old wording and move on — we keep the correction date and the reason. Withdrawn measures, such as using the Indexing API for general pages, stay on the record as they are.

How to read this log

Each measure is recorded under four points.

Implemented: What we changed.
Observed: What happened, in which data.
Current assessment: Continuing, on hold, or withdrawn.
Limits: Why it cannot be stated as an outcome.

Volume of work and number of pages published are productivity information. Search traffic, mentions in AI, and inquiries are outcome information. We keep the two apart.

The record of measures

The date we carried out a measure and the date Google updated its official position are treated as separate things. Withdrawn measures are not deleted; they stay on the record together with the correction and what we changed to prevent a repeat.

Measure	What we did	What we observed or corrected	Current assessment and limits
1. Fixing redirects from old URLs	We fixed redirects that were sending old paths to blog URLs that did not exist, and health-checked the canonical URLs.	The target URLs now reach the intended article instead of returning a 404.	Complete as a fix for a technical fault. Any improvement in search traffic cannot be attributed to this fix alone. Limit: it takes time for search engines to re-crawl the old URLs and consolidate how they are evaluated.
2. Reviewing structured data	We reviewed places where Product structured data had been used for services that are not physical goods, and changed them to a type that matches what the page actually is. We also checked that the displayed content and the structured data agree.	The relevant warnings in Search Console were resolved.	Effective as an improvement in technical consistency. It is not, however, special schema for appearing in generative AI features. Limit: resolving a warning, and any effect on ranking, citation in AI answers, or inquiries, have to be confirmed separately.
3. Article-specific OGP	We set a per-article image and article attributes for each blog post.	Sharing on social platforms or in chat now shows the article-specific image and title.	An improvement that helps people understand a link and decide whether to click it. Limit: effects on social traffic or branded search depend on posting frequency and topic.
4. Adding internal links	We added paths from service pages to related technical articles, and from articles to related services and related articles.	Orphan pages decreased, and the structure now allows movement to related pages.	We continue this as a baseline improvement needed both for understanding the site and for crawling. Limit: adding links alone is not enough; we check periodically that the context and the destination match.
5. Improving titles and descriptions	We identified pages with impressions but few clicks, and changed titles and descriptions to match search intent.	Some pages are flat, and we have not confirmed a change that could be judged a clear improvement.	Under continued observation. Rather than repeating title changes, we check that search intent and body content agree. Limit: Google may rewrite the text shown in search results, and competitors, seasonality, and ranking movement also have an effect.
6. Installing llms.txt	We installed an llms.txt summarizing the site overview and the main pages in Markdown.	We confirmed the file is published and can be retrieved. We have not confirmed a causal link to any increase in citation or recommendation in AI search.	An experimental supporting file. We do not treat it as a required measure for Google Search. Limit: not every AI service reads it, and the way it is processed is not standardized.
7. Sending general URLs to the Indexing API	In April 2026 we submitted general site URLs to the Indexing API.	Correction: this use was outside Google’s official scope. The Indexing API applies only to pages with JobPosting, or with BroadcastEvent inside VideoObject. Source: developers.google.com/search/apis/indexing-api/v3/using-api	Withdrawn, and the processing stopped. We do not recommend it as a way to push general pages into the index. To prevent a repeat, we added a gate that checks the target content, the purpose, the official documentation, and the stop conditions before any API is used.
8. Setting a policy for AI crawlers	We separated crawlers used for search and reference, for training, and for user-triggered fetches, and reviewed our robots.txt policy.	We made the settings and the server logs checkable. Source: developers.openai.com/api/docs/bots	In line with our policy of having public information discovered in AI search, crawlers used for search and reference are allowed by default. Training use is judged separately. Limit: allowing a crawler is a setting that secures the possibility of retrieval; it does not guarantee appearance in a search answer.
9. Observing questions in AI search	We built question sets divided into the customer’s problem, comparison, and purchase stages, and observed mentions, citation, and accuracy in outside AI.	We confirmed that results change with the wording of the question and with when it is run.	We dropped single-shot judgments and moved to repeated observation under fixed conditions. Limit: model updates and changes in search results mean the causal link with a measure cannot be fully isolated.
10. Measuring Google’s generative AI features	We changed the design so that the generative AI performance report in Search Console is reviewed separately from ordinary web search, citation by outside AI, and AI referrals.	—	Our policy is not to mix Google-internal and outside AI into a single visibility score. Limit: the screens provided and the scope they cover may change, so we record the date we checked.

We can start with you from the recording and observation design covered in this log. A free diagnosis based on public information shows how things currently look.

Check how search and AI describe your site

What cannot be claimed as cause (separating hypothesis from fact)

As the record above repeats, for most measures we have not proven the causal claim that "we did this, so AI carried us" or "so our ranking rose." Refusing to leave that ambiguous is the main point of this log.

What can be stated as fact: The day a measure was carried out, what was carried out, and the result we could observe (for example, the verifiable fact that a new page was indexed).
What stays a hypothesis: An early qualitative impression — a sense that "our company name seems to come up more in AI answers" — is not data showing cause; it is still a hypothesis. The sample is small, and AI answers vary by model, search feature, region, and point in time, so one or two observations cannot settle it.
The premise Google states officially: Official guidance says there is no special requirement for appearing in generative AI search beyond being indexed and eligible for snippet display. A direct causal link between an individual measure and AI visibility is not stated officially either.

This log is therefore not a collection of success stories. Being able to speak about cause would require fixing an observation period, repeating measurement under the same conditions, and building up a larger sample. We are partway through that. The terminology itself is sorted out separately.

What Are AIO, GEO, and LLMO?

Measures where we could not confirm an effect

Alongside the successes, we keep on record the measures where we could not confirm an effect.

Effect of llms.txt on Google Search: We installed it, but could not confirm an effect on display or traffic in Google Search. Google states officially that there is no need to create a new machine-readable file, AI text file, markup, or Markdown in order to appear in Google Search, and does not list it as a ranking factor.
FAQ structured data producing a rich result: Google ended the display of FAQ rich results in search results on May 7, 2026. Keeping FAQ structured data in place is not a problem, but a rich result in search listings can no longer be expected. We correct the earlier version, which described FAQPage as "a measure with a display effect." The FAQ in this article is published as visible HTML only.
Title and description changes raising CTR: Click-through rate on some of the pages we rewrote has stayed flat, and no improvement has been confirmed so far. Because reflection and re-evaluation take time, we do not call it a failure; we keep it under continued observation.

Hiding these would turn a practice log into promotional copy. A record of what did not work is exactly what informs the next decision.

What we verify next

These are the points we are still observing. When results come in, we will add them to this log with the observation period fixed.

Whether we enter the candidate set for customers’ non-branded questions.
Whether citation and reflection in the answer agree.
Whether AI referrals proceed to the main CTA.
Whether branded search is increasing.
Whether improvement proposals reach implementation and publication.
Whether the content of customer inquiries has changed.

A checklist for working from facts

Items to confirm if you do the same thing for your own site on a factual basis.

Are you recording "implemented / observed / current assessment / limits" separately for each measure?
Are you keeping "the day you carried out the work" and "the day Google updated its official position" from being conflated?
Are you distinguishing, in the writing itself, "what you observed as fact" from "an impression or hypothesis"?
For AI visibility, are you avoiding stating cause for an individual measure? (Google states officially that special schema, a special writing style, and AI text files are not needed.)
Are you avoiding adding FAQ structured data in the expectation of a rich-result display effect? (Display ended on May 7, 2026.)
Before using an API or an outside service, are you checking the target content, the purpose, the official documentation, and the stop conditions?
Is baseline SEO in place — crawling, indexing, snippet eligibility, internal links, mobile display?
Are you keeping measures with no confirmed effect on the record rather than hiding them?
Have you fixed the observation period, and are you repeating the same observation under the same conditions?

Frequently asked questions

Should we install llms.txt?: There is no need to install it in the expectation of an effect on Google Search. Google’s official guide states plainly that AI text files and the like do not need to be created in order to appear in Google Search. llms.txt is not listed there as a ranking factor, and we could not confirm an effect after installing it either. It does no harm, but the grounds for spending limited effort on it are thin.
Is there any point in adding FAQ structured data?: A rich result in search listings can no longer be expected: Google ended the display of FAQ rich results on May 7, 2026. Keeping the structured data itself is not a problem. Our current assessment simply does not treat it as "a measure that increases display." In this article we do not attach FAQPage structured data, and keep the FAQ as visible HTML only.
Can the Indexing API push our pages into the index?: For general pages it is out of scope. The Indexing API applies only to pages with JobPosting, or with BroadcastEvent inside VideoObject. We were submitting general URLs in April 2026; because that use was out of scope, we withdrew it and stopped the processing.
Why are there so few outcome figures in this log?: netsujo.jp is a young site, traffic is still very small, and results that could be spoken of as cause have not accumulated. We do not write that results exist when they do not. We record verifiable facts (for example, a new page being indexed) separately from impressions that remain hypotheses, and we are still observing.

Corrections and update history

In a practice log, what we judged correct at the time of writing can later be overturned. We keep that history too.

Date	What changed
April 2026	We submitted general site URLs to the Indexing API. This use was withdrawn on July 23 of the same year (measure 7).
June 2026	We put structured data (JSON-LD) in order, installed llms.txt and llms-full.txt, unified our notation, adopted a conclusion-first structure, designed hub-and-spoke internal links, allowed AI crawlers, reviewed titles and descriptions, and recorded a measurement baseline. The record of what was carried out is carried over from the earlier version as it was.
July 23, 2026	We withdrew the use of the Indexing API for general pages, and added a note that llms.txt, ai-plugin.json, and similar files are not treated as requirements for appearing in Google’s AI features.
August 2026	We added the Google generative AI performance report to the measurement design, separating Google-internal, outside AI, traffic, and conversion. We began consolidating older articles and adding correction paths.

Earlier wording	Correction
Presented llms.txt as an effective measure	Corrected. Google states officially that AI text files and the like do not need to be created, and we could not confirm an effect on Google Search ourselves
Claimed a rich-result display effect from FAQPage	Corrected. Google ended the display of FAQ rich results on May 7, 2026, and we removed the FAQPage structured data from this article
Implied a causal link between structured data and AI visibility	Corrected. The causal link has not been observed
Presented early qualitative impressions as results	Now stated explicitly as hypotheses
Listed only successful measures	Added the measures where we could not confirm an effect
Mixed the date we implemented something with the date Google updated its official position	Now treated as two separate things

Why we publish this log

In SEO and AI search work, the name of a measure tends to run ahead, and simply implementing something can be described as if it had produced a result.

We record the measures where we could not confirm an effect, the measures that were wrong, and the measures we are holding judgment on. What continuing web improvement needs is not a stack of success stories but an operation that can be corrected.

Diagnosis from public information, analysis in GA4 and Search Console, and the design of baseline SEO and measurement. You can start from the free diagnosis.

Check first, for free, how search and AI describe your company