What 5,500 Shipped Fixes Taught Us About Getting Cited by AI

Most advice about AI search is borrowed. Someone reads a study, summarises it, and passes it along. This article is different: every number in it comes from our own production system, the one that finds, ships, and verifies SEO fixes on real Australian business websites every week.

Between March and June 2026 our SEO Agent proposed 6,888 improvements across nine websites (seven clients plus our own two). A senior SEO approved 5,554 of them, our WordPress plugin applied them, and we re-checked roughly 2,200 on the live page to confirm they actually rendered. Alongside that, we ran more than 2,400 live citation checks across ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude, and 611 scheduled audit runs.

Here is what that data says, including the parts that are inconvenient.

How we collected this (and the honest caveats)

Every claim below comes from our own database, pulled on 13 June 2026. Three caveats before you quote any of it. First, nine websites is a meaningful operational dataset but not a market study; treat the patterns as directional. Second, our citation checks use the engines' APIs as proxies for the consumer apps, and results can differ from what a person sees in the app. Third, the audit-run data covers sites we monitor, which means some had already been partly fixed; the true failure rates for untouched websites are likely worse, not better.

No numbers in this article are estimates or industry borrowings. If we could not trace a figure to a record, it is not here.

What stops most websites from being cited?

The single biggest blocker is not technical. Across 611 audit runs, 88 per cent of pages failed our source-citation check: they cite no external references at all. And 43 per cent lacked any statistical content worth quoting.

That matters because AI engines cite pages the way journalists cite sources. A page that references government data, industry bodies, or named studies looks like evidence. A page of unsourced sales copy looks like an opinion, and engines treat it accordingly.

The rest of the failure list is more conventional: 29 per cent of audited pages were missing FAQ schema, 25 per cent had H1 structure problems, and 24 per cent had meta descriptions that were missing, empty, or truncated. All fixable. All routinely unfixed.

Which AI engine actually cites businesses?

Google AI Overviews, by a wide margin. On queries we actively optimise for, our 770 targeted checks found businesses cited in AI Overviews 33 per cent of the time, Perplexity 28 per cent, ChatGPT 22 per cent, and Gemini 20 per cent. Claude cited in under 2 per cent of checks.

Zoom out to the broader weekly tracking set (1,650 checks across 71 queries, including many we have not optimised yet) and the numbers drop hard: AI Overviews 11.5 per cent, Perplexity 7.6 per cent, and every other engine under 2 per cent.

Two lessons. One: Google AI Overviews and Perplexity are where citations happen first, because both retrieve live web results every time they answer. Two: for most queries, most businesses are simply not cited anywhere. If you run a check on your own business and find nothing, that is the normal starting point, not a special failure.

What does the work actually look like?

People imagine AI search optimisation as a mysterious new discipline. The change log says otherwise. The fixes that shipped most often were answer blocks (666 applied), meta descriptions (548), page titles (541), entity schema (513), FAQ schema (432), and llms.txt files (404). We also corrected robots.txt rules that were blocking AI crawlers on 215 occasions, usually by accident, where a business trying to opt out of AI training had also locked out the crawlers that fetch pages for AI answers.

None of this is glamorous. It is hundreds of small, structural changes that make a website readable, quotable, and attributable. The compounding effect is the point.

Why do recommendations pile up while nothing ships?

This is the finding we care most about, because it explains why so many businesses pay for SEO and see nothing change.

When a fix can be applied automatically after approval, it ships. Meta descriptions: 592 proposed, 548 applied. Entity schema: 539 proposed, 513 applied. But when a fix needs a human to write or restructure content, the funnel collapses. Content suggestions: 300 proposed, 8 applied. Heading rewrites: 174 proposed, 9 applied.

Same system. Same clients. Same approval process. The difference is execution. Advice queues; shipped changes compound. Every monitoring dashboard and audit tool on the market produces the first column. The entire reason we built our agent the way we did, with a senior SEO approving and a plugin applying, is the second column.

Do the fixes survive contact with the real website?

Not always, which is why we check. After our plugin applies a change, our system fetches the live page and confirms the change actually renders. Around 2,200 shipped changes have been re-verified on the live page so far. The failures are instructive: page-builder layouts that ignore content updates, caching layers serving stale HTML, and themes that overwrite meta tags. If your agency reports "done" without looking at the live page, some of that work does not exist.

Does any of this produce actual citations?

Yes, with patience and no guarantees. Chatime Franchising Australia started with us in early April 2026. Within nine weeks their audit score went from 83 to 93 and they were being cited in Google AI Overviews for tracked franchise queries, plus a Perplexity citation. That is what early movement looks like: specific queries, one engine first, then spread.

The longer arc looks like Pentana Stanton Lawyers, whose organic search sessions grew from 4,056 in Q1 2023 to 27,510 in Q1 2026, up 578 per cent, with monthly Google clicks up roughly eightfold. Search visibility built this way is the raw material AI engines draw on. The full numbers, with sources, are on our results page.

We will not promise you a citation by a certain date. AI engines are third-party systems and nobody controls them. What we can promise is the work, the evidence trail, and the re-checking.

What should you do with this?

If you take nothing else from our data, take the failure list, because it doubles as a priority list:

1. /Add sources to your key pages. Link to government data, industry bodies, and named studies. This was the most-failed check in our entire dataset.
2. /Publish real numbers. Pages with statistics get quoted; pages without them get skipped.
3. /Answer questions directly. Question-shaped headings with a straight answer in the first sentence, backed by FAQ schema that matches what is visible on the page.
4. /Check your robots.txt. Make sure you are not blocking AI citation crawlers by accident. It was one of the most common fixes we shipped.
5. /Verify everything on the live page. Whoever does your SEO, ask them to show you the change rendering on the live site, not a report saying it was done.

Or let the system that produced this dataset do it for you. The free AI visibility audit runs the same six-category checks this research is built on and shows you exactly where you stand. If the grade stings, that is the data being honest with you, and it is the same honesty we bring to the work.

About the author

Jayson Munday

Founder - AEO & SEO Strategist

20+ Years in SEO & Digital Marketing22 years in practice

Founder of Brain Buddy AI with over 20 years in search marketing. Jayson identified the AI search revolution early and built one of Australia's first managed SEO, AEO, and GEO service to help businesses get found by every AI engine.

SEOAEOGEOContent StrategyLead Generation

More from Jayson →LinkedInWebsite