We use AI through most of our audit work. It has cut the time on a full technical audit from about two weeks to four days. It has also produced confidently wrong output that would have cost a client months if we had shipped it. Both things are true, and knowing which is which is most of the skill.
Where it genuinely helps
Pattern extraction across large crawls. A 200,000-URL crawl export is not something a human reads. Feeding structured crawl data through a model to surface clusters — templates with missing canonicals, page types with systematically thin content, parameter patterns generating duplicates — takes minutes and finds things a manual pass misses.
Log file summarisation. Server logs tell you what Googlebot actually does rather than what you assume. Turning a month of raw logs into “these six directories absorb 70% of crawl requests and produce 2% of traffic” is exactly the kind of aggregation AI is good at.
Content gap clustering. Given a competitor keyword export, grouping thousands of terms into coherent topics is fast and reliable. We still review the groupings, but the first pass is close.
First-draft schema and briefs. Structured data markup and content briefs are formulaic. Generating them and then correcting is faster than writing from scratch.
Where it fails, and fails quietly
Search intent judgement. A model will tell you a keyword is informational. Look at the actual SERP and it is nine product pages. Intent is determined by what Google currently rewards, not by the grammar of the query. We check the live results every time.
Competitive context. AI does not know that the site ranking first is a government portal you will never outrank, or that a competitor bought their backlink profile and is one update away from losing it. That judgement comes from having watched the market.
Thai language nuance. Tokenisation of Thai is still weak in most models. Keyword variants get merged that should be separate, and genuine synonyms get split. For Thai-language work we treat AI output as a starting list to be manually checked, never as the list.
Anything numeric it was not given. Ask a model for search volumes and it will produce plausible numbers from nowhere. Every figure in our audits comes from Search Console, Ahrefs, or GA — never from a model.
The rule we work to
AI handles volume; people handle judgement. If a task is “process a lot of data into a shape”, it goes to the tool. If a task is “decide what matters and in what order”, it stays with us.
The reason we are strict about this is that AI errors do not look like errors. A wrong crawl summary looks exactly like a right one. The only defence is knowing which outputs you are allowed to trust without checking, and we keep that list short.
