What an AI Search Audit Should Change in Your Marketing Plan

A useful AI search audit changes at least one decision about budget, technical work, content, source correction or measurement. If it only produces a score, it has told you where you stand and left the plan exactly as it was.

To show what that looks like, we audited our own site, search.agency, in early November 2026 and took three findings all the way to a decision. Search Agency wrote this article and sells AI search audits, so we have a commercial interest in the conclusions. We have kept the raw answers, scripts and checks, and every number below comes from them.

The decisions an audit should change

Most marketing plans have five places where an AI search audit can move money or time. Each one is a decision a director already owns.

• Budget allocation, meaning whether AI search gets money this quarter and how much.
• Technical sequencing, meaning which fixes developers ship first.
• Content priorities, meaning which buyer questions get a page and in which language.
• Source correction, meaning which wrong or conflicting facts about you get fixed, and where.
• Measurement, meaning what you track so you know whether any of it worked.

Read an audit against that list. A finding that changes none of these five is background reading. A finding that changes one of them should come with the evidence, the person who will act on it and a way to tell whether it worked.

The bar we used for each finding below has seven parts. We record the observation, list the explanations that could account for it, check them, make a decision, name an owner and a dependency, and set a success criterion. A missing mention in an AI answer can justify an investigation. It does not tell you the cause, and the cause is what decides where the money goes.

What the audit actually examined

This was a small, single-surface study. We collected answers only from Google AI Mode, through a third-party collection service called SearchAPI, with the location set to Jakarta. That is a different thing from a person typing into the Google app, and the results should be read as an AI Mode sample collected by API.

The buyer questions were the ones a marketing director in Indonesia might ask when looking for help, such as the best AI search agency in Indonesia, a GEO agency in Jakarta, and what SEO costs per month. GEO means generative engine optimization, the work of getting a brand named and cited in AI answers. The repeated questions are the same buyer asking twice, so they show stability and add no extra buyers.

Twenty answers is enough to find problems worth checking and too few to estimate a market share. We treat every rate below as a description of this panel on these two days.

The first finding is an unindexed page

The observation started with a suspicion we had to test. Search Agency appeared in very few commercial answers, and a common explanation for that is a technical block that stops crawlers reading the site.

We checked that first. A technical block would make every content decision pointless. The robots.txt file allows every crawler and names a dozen AI crawlers explicitly with an allow rule. Five key pages returned a normal 200 response to Googlebot, GPTBot, PerplexityBot, ClaudeBot and Bingbot. The raw HTML already contains the visible text, so nothing important depends on JavaScript. Structured data and self-referencing canonical tags were present on all seven key pages we tested.

Search Console then showed a smaller and real issue. We checked eight key URLs, and seven were indexed. The eighth was a blog post on evaluating SEO work, published on 8 September 2026 and listed in the sitemap.

https://search.agency/blog/evaluate-seo-work. Discovered, currently not indexed. URL inspection result, URL is unknown to Google.

There were three competing explanations. The page might be blocked or point its canonical elsewhere. It might be poorly linked internally. It might simply be low in Google's crawl queue. The first was ruled out by the page itself. For the second, we counted internal links across all 215 sitemap URLs. Then we checked the ten blog posts with the fewest links. Nine of those ten were indexed, so weak linking does not explain the gap on its own. Across the 18 URLs we checked, two were not indexed.

The decision is to request indexing for the two pages and add contextual links from related posts. No money goes to a technical program. The owner is the SEO lead, and the dependency is a developer or editor who can add the links. Success means both URLs show as indexed in Search Console within four weeks.

What would change this decision? If a wider index check found more than a handful of recent posts unindexed, the problem would be systemic, and a crawl and site-architecture review would move to the top of the roadmap.

The second finding is a missing buyer question

Across the 16 commercial answers, Search Agency was named in 2 and its domain was cited in the same 2. Every commercial answer cited at least one source, so the share of cited answers that included search.agency was also 2 of 16. Both mentions came from one question, "best AI search agency in Indonesia", which named us in both runs.

Counting each agency once per answer, the 16 answers produced 36 agency mentions, and Search Agency had 2 of them. Our Indonesian sister site, aisearch.id, appeared in 4 answers, which lifts the group to 3 of the 12 commercial questions.

In Indonesia, several digital marketing agencies specialize in GEO, entity recognition, and AI visibility.

The cited sources tell you where these answers come from. They were mostly third-party "best agency" lists, directory pages and other agencies' service pages, many written in Bahasa Indonesia. We fetched two of the lists that AI Mode cited, and neither mentions Search Agency.

We considered four explanations. Our pages might not use the buyer's words, but our titles already use GEO, AEO and AI search. The site might be invisible to crawlers, which finding one ruled out. The answers might be unstable, and one question did change almost completely between runs. Or the sources AI Mode leans on for these questions might simply not include us, which the fetched lists support.

The decision is a bounded content and outreach test. We will publish one Bahasa Indonesia page on aisearch.id that explains what an AI visibility audit checks. Bahasa service pages were what AI Mode cited for that question. We will also ask the editors of the two cited lists to consider us for review, knowing they owe us nothing. The owner is the content lead, and the dependency is indexing of the new page. Success means a group mention or citation in at least one of three runs of the audit and GEO-service questions, eight weeks after indexing.

What would change this decision? If repeated runs showed the cited sources rotating heavily each week, a single page would be too small a bet. The test would then need more questions and more runs before any budget moves.

The third finding is an accuracy problem on our own pages

We asked four questions about Search Agency itself and extracted 25 factual claims from the answers. We checked 23 of them against our own pages as they stood on 2 November 2026. Twenty-two were correct, covering the founder, the founding date, the office address and the published price tiers. One was wrong, one was ambiguous and one was an opinion that cannot be checked.

It is structured primarily as an exploratory diagnostic rather than a do-it-yourself roadmap.

The more useful discovery came while we were checking the correct answers. Two sections of our audit page say the benchmark covers 3 to 5 AI engines. The question-and-answer section on the same page says 7 to 10, and so does the structured data behind it. AI Mode repeated 3 to 5 twice, so it was right by picking one of two conflicting official statements. A different run could pick the other.

The fourth question used only our name, "Search Agency Indonesia pricing". The answer gave generic Indonesian SEO price ranges and never named us, while the three questions that included the domain resolved correctly. Search Console points the same way. Over 90 days the query "search agency" earned 1,731 impressions at an average position of 31.2, and some impressions came from searches for an unrelated company in India with a similar name.

We cannot see the whole retrieval path inside AI Mode, so we will not claim the conflicting engine count caused anything. What we can say is that two of our own pages disagree, and that our name is a generic phrase.

The decision is to fix the audit page first, because it costs an afternoon. One set of numbers for engines, competitors, timeline and deliverables goes on the audit, methodology and pricing pages and in the structured data. The owner is the founder as offer owner, and the dependency is the web developer. Success means every page states the same values and a rerun of the audit question repeats them. Work on the name collision waits for evidence, as the next sections explain.

What would change this decision? If name-only questions start to return wrong facts, such as another company's prices attributed to us, disambiguation work would move up the list.

From findings to a roadmap

Evidence strength and business priority are different scores, and a roadmap needs both. The page inconsistency has strong evidence and low cost, so it goes first. The buyer-question gap matters most commercially, but its evidence comes from 16 answers on one surface, so it gets a bounded test before a budget line.

The sequence follows dependencies. The offer page has to be settled before anyone writes a Bahasa page describing the audit. The new page has to be indexed before its eight-week clock starts.

The rerun uses the same questions, the same surface and the same location. Changing the panel halfway would make the comparison meaningless, which is also why our published methodology freezes the prompt set during an engagement.

What should not change yet

The finding that did not justify investment was AI crawler access. A common audit recommendation is to rewrite robots.txt rules, add an llms.txt file or rebuild pages for server-side rendering. On search.agency all of that already works, so the spend would buy nothing.

The name collision is the example of an unresolved finding. One name-only answer failed to resolve, and Search Console shows impressions from another company's searches. That could justify entity work later. It could equally be harmless, because buyers who already know us tend to type the domain or the founder's name. One answer is too thin to tell which, so the honest status is unresolved, with a monthly check.

Budget for AI search as a whole also stays where it is. Nothing in these 20 answers tells us how many buyers ask these questions in AI Mode, or whether an answer changed who they contacted. Search Console shows 4,265 impressions over 90 days for "ai search agency" at an average position of 3.9, so classic search still carries real demand. A budget increase should wait for the week-12 rerun and a look at enquiries. If you want the groundwork on that judgment, our guide to what your business needs in AI search covers it.

How to evaluate the handoff

When an audit lands on your desk, five questions tell you whether it is ready to act on.

• Does it state the surface, the collection method, the dates and the number of answers actually collected?
• Can you open the raw answer behind each headline number?
• Does each finding list other explanations and show how they were checked?
• Does each recommendation name an owner, a dependency and a success criterion?
• Does it say clearly what should not change, and what remains unresolved?

An audit that passes all five can change a plan. One that fails the second or third is still a list of opinions, however polished the scorecard. Our article on how to evaluate SEO work applies the same test to ongoing retainers.

A free benchmark is a starting diagnosis. It does not include implementation, unlimited technical consulting or a guarantee of citations, and nobody can promise placement inside an AI answer. If you want this kind of evidence for your own brand, request a benchmark audit with a decision-ready roadmap.