Chapter 2 of 5 · AI and SEO — Using AI Well, and Understanding AI Search
Where AI genuinely helps with SEO, and where it makes things up
AI is good at drafting titles, descriptions, structured data and summaries of your Search Console data, but do not trust the numbers, quotes or Google rules it gives you unchecked. With the lessons this site learned from information that sounded right and was wrong.
The last chapter concluded that Google does not care who wrote a page, only whether it is useful and accurate. This chapter therefore splits SEO work into two groups: tasks AI does well, and tasks you must check every time because AI can be confidently wrong.
2.1 What AI does well
Google itself says generative AI is useful for researching a topic and adding structure to your own content. Tasks that fit:
| Task | How to use AI | What to check |
|---|---|---|
| Titles and descriptions | Ask for several versions, then pick the one with words people really search | The words must come from Search Console’s Queries tab, not from the AI |
| Article outline | Have it propose sections from the material you have | Cut any section it adds without supporting material |
| Structured data (JSON-LD) | Have it draft from the page content | Test with the Rich Results Test; the data must match what the page shows |
| Image alt text | Have it draft from a description of the image | It must match the actual image |
| Translating a draft | A fast first draft | A speaker of that language must review it |
| Summarising a Search Console export | Have it group rows or find patterns | Open the pages or rows it refers to, every time |
What these share: you own the data. The AI only reshapes or rewords it, and the result is easy to check against the real thing.
2.2 Where AI is confidently wrong
| What the AI tells you | The problem | What to do |
|---|---|---|
| Search volumes or statistics | If the tool is not connected to a real data source, the number is a guess | Ask where the number comes from; if there is no answer, don’t use it |
| Quotes and source links | The wording may not match the original; the link may not exist | Open every link and find the wording on the real page |
| Rules from Google or other platforms | Rules change often; the AI’s knowledge may be older than the current rule | Check the official documentation and its last-updated date |
| Words Thai users search | The same thing can be written several ways; the AI may pick one nobody types | Look at real queries in Search Console |
| Code or markup | Looks right but breaks the spec | Test it with a validator |
An example of a rule that changed this year: since 31 August 2026, Search Console lets every site choose whether it appears in AI Overviews, AI Mode and AI features in Discover (chapter 4). An AI trained before then knows nothing about it.
2.3 A lesson from KoishiAI: summaries that sounded right and were wrong
While building this site we hit information that “looked right” but was wrong several times, and each time the fix was the same: open the primary page and read it.
The rule for reading numbers. Our baht-text tool page once said, following summaries from search results, that a final 1 may be read either “เอ็ด” or “หนึ่ง” under the Royal Institute’s rules. The Royal Institute’s site would not load at the time. When it did and we read it ourselves, its knowledge note says that under Thai rules a final 1 is read “เอ็ด” only. We corrected the tool page.
A tool that no longer exists. We had recommended LINE’s link-preview cache tool, as several sources did. When we checked, its hostname no longer resolved.
Our own drafts. Our tutorials and the articles we researched ourselves go through a fact-check separate from the writer: another AI, which did not write the draft, opens the primary pages and checks each claim. Every set had errors in its first draft — a wrong explanation of a cause, a missed finding, sentences that went beyond the evidence.
What we learned: the most dangerous wrong information is not the kind that looks odd, but the kind that sounds reasonable and matches what you expected.
2.4 Researching with AI without getting caught
- Let the AI point the way, then read the source yourself. Ask which page the information comes from, and open it.
- Check dates. Google Search Central pages show a last-updated date at the bottom. If two pages disagree, trust the newer one and note the difference.
- Separate facts from interpretation. “Google says X” and “we think X because Y” must be distinguishable to the reader.
- Have someone who did not write it check it, whether the writer was a person or an AI.
2.5 After this chapter
- Pick 3 pages written with AI help, underline every number, quote and rule, and find the source for each.
- Where you cannot find a source, delete the claim or mark it clearly as opinion.
- Change the words in your titles to match real queries from Search Console.
The next chapter puts this together into a system: AI writing with sources, checks and a reviewer.
Related tools
- JSON-LD builder & checker — draft structured data and check it against Google’s rules
- Link preview checker — see how your titles and descriptions look when shared
Sources
- Google Search Central — Using generative AI content
- Google Search Central — Documentation updates
- Search Console Help — Search generative AI control
- Royal Institute (Office of the Royal Society) — หนึ่ง or เอ็ด (Thai)
- Royal Institute (Office of the Royal Society) — Reading numbers of two or more digits (Thai)