How to rank in AI search: what we learned taking 13 AI answers apart

Nobody types “plumber melbourne” into ChatGPT. They ask “who is the best plumber in Melbourne and what will it cost”. The engine writes one answer and names three to five businesses. There is no page two.

We wanted to know where those names come from. So instead of reading other people’s studies, we ran the test ourselves. Thirteen real buying questions, one per client type, with every page the engine read and every sentence it quoted written down. This article is that data, what it means for how to rank in AI search, and the checklist we now run for every client.

Key findings

What we tested

On 10 and 11 September 2026 we ran 13 buying questions matching the businesses we work with, including plumbers, solar, home builders, employment lawyers, mortgage brokers, psychologists, ATAR tutoring and industrial electric motors. Three question types: “best X in city”, “how much does X cost” and “how to choose X”.

For each question we ran the same retrieve-then-write loop the AI engines use: search, read the top pages, write an answer with sources. We logged every URL read, which of three layers it belonged to (directory, third-party list, or the business’s own site), which businesses were named, and the exact sentence lifted from each page. Then we checked the pattern against the 2026 studies from Ahrefs, SE Ranking, Seer Interactive and SOCi. It held.

One limit worth stating. We ran the loop with a web search tool, not inside the ChatGPT or Gemini apps, so the businesses named could differ from what a customer sees on a given day. The source pattern, which is the part that decides what to do next, matched the published research.

How an AI answer is built

Four steps. The same on Google AI Mode, ChatGPT and Perplexity.

  1. It takes your question. “Best plumber in Melbourne and what will it cost.”
  2. It splits the question into five to ten smaller searches. Best plumber Melbourne, plumber reviews, call-out fee Melbourne, licensed plumber VIC, and so on. Google calls this query fan-out and named it in the AI Mode announcement on 5 March 2025. The mechanic sits in Google’s own patent filing.
  3. It reads the top pages for each search and pulls a few sentences from each. It does not read the whole page.
  4. It writes one answer and names its sources. Usually about four.

ChatGPT runs the same loop on the Bing index. Seer Interactive found 87% of ChatGPT search citations match Bing’s top results. Google’s matched only 56%.

One consequence matters more than the rest. The engine reads sentences, not pages. One good sentence beats a long page.

Finding 1: the names come from directories and lists

Take “best plumber in Melbourne”, run on 11 September 2026. Seven pages were read before the answer was written. Four were directories: Oneflare, Localsearch, ProductReview and Yelp. One was a “7 best plumbers in Melbourne” article, written by a plumber who appears on its own list. Two were business websites, read only for facts.

The answer named five plumbers. Four of the five names came from the directories and the list.

The same three layers showed up in all 13 questions, from plumbers to employment lawyers.

  1. Directories and review sites. They hand the engine a ready-made ranked list with rating, review count, years in business and price band. Oneflare and ProductReview for trades, Clutch for agencies, Canstar for builders, Doyle’s Guide for lawyers.
  2. “Top N in city” articles. Mostly written by competitors who put themselves first. A name that appears in three lists reads as the consensus.
  3. The business’s own website. Read last, and only for facts to quote.

Your own site decides what gets said about you. It does not decide whether you get named. That is the single biggest change from classic SEO.

Finding 2: AI search quotes numbers, never adjectives

The quoted line almost always came from the title tag, the meta description or the first paragraph. One Melbourne plumber’s title tag reads “Plumber Melbourne: 3,000+ 5-Star Reviews”, and that is the line the answer used. Two of our own clients were quoted word for word from their meta descriptions.

Other lines that got quoted across the 13 questions: “4.9 stars from 2,000+ reviews”, “over 35 years”, “on site in 90 minutes”, “HIA award 2025”, “35+ lenders”, “since 1994”, “120+ team, largest in Australia”.

Lines that were never quoted, not once in 13 answers: “trusted local plumber”, “reliable, friendly service”, “quality workmanship”, “we go above and beyond”.

A model can quote a fact. It cannot quote a feeling.

Finding 3: the cost question goes to whoever wrote the page

Second test: “how much does a plumber cost in Melbourne”. No directory this time. Eight pages were read and six were plumbers who had written their own cost guide. The answer quoted their numbers: $90 to $180 an hour, $80 to $150 call-out, $200 to $450 for a small job.

One page, one question, and you are in the answer. Every business has ten questions like this with no page. That is the content plan.

Finding 4: the locale trap for Australian businesses

We asked for the “best gutter company Melbourne”. The answer came back with Melbourne, Florida: Angi, Yelp US and HomeAdvisor. Adding “Australia” to the question fixed it. Your customer will not add it.

The page has to say where it is, in words the engine can read. City, state and country in the title and the first paragraph. Street address written out on the page, not only in a footer image. Service suburbs listed by name. LocalBusiness schema with addressRegion and addressCountry filled in. “Melbourne” is not enough. Say Melbourne, Victoria, Australia.

Why rankings alone do not get you into AI answers

Our test says the names come from other people’s pages. The big studies say the same thing with bigger numbers.

Ahrefs checked 75,000 brands against AI Overviews. Branded web mentions correlated with AI presence at 0.66. Backlinks scored 0.22. YouTube mentions scored 0.74, higher than anything else measured. Brands in the bottom half for web mentions were, in Ahrefs’ words, “essentially invisible to AI”.

A separate Ahrefs study of 15,000 prompts found only 12% of URLs cited by ChatGPT, Gemini and Copilot rank in Google’s top 10 for the same prompt.

Page one on Google is the entry ticket. Mentions elsewhere get you named.

How to get your name into AI answers

This is the part most SEO plans skip, and it is where the names come from. SOCi’s 2026 Local Visibility Index covered 350,000 business locations. ChatGPT recommended 1.2% of them. The Google 3-pack showed 35.9%. The gap is profile work.

How to get your page quoted

Once the name is in the answer, the site decides what gets said. Every item here aims the site at the sentences the engine actually copies.

How to track AI search visibility

Rank tracking for AI answers works like keyword tracking: a fixed list of prompts, checked weekly, per engine. We use 15 prompts per business. Eight unbranded (“best plumber in Melbourne”, “blocked drain plumber Melbourne”), three comparison (“how to choose a plumber”), two cost (“plumber call-out fee Melbourne”), and two branded, tracked separately so they do not inflate the score. The prompts come from the business’s own Search Console question queries, so they are things people actually ask.

Four numbers per engine. Visibility, the share of runs where the business is named. Citation rate, the share of runs where its own page is cited. Share of voice against up to five named competitors. And the prompts that moved most, with the sentence quoted. Put the question-shaped queries from Search Console next to them so AI and Google sit side by side.

The first 30 days is baseline. AI answers change run to run, so judge the prompt set as a group, never a single prompt in a single week.

Every prompt then lands in one of three results, and each result has one action. Not named and no page on the site: write the page. A competitor named via a directory or a list: get onto that directory or list. Named, but the quoted fact is weak: fix the proof line on that page. That is the content plan, and it rewrites itself every month.

FAQ: ranking in AI search

Is ranking in AI search the same as SEO?

It starts from SEO and adds a layer. Page one on Google or Bing gets a page read. Mentions on directories, lists and third-party sites get the business named. Classic SEO covers the first part only.

Does schema markup get a page cited by AI?

Not on its own. Two separate studies found pages that added schema saw little or no lift in citations. Put the proof line in the title and first paragraph first, then add the markup.

Which AI engine should an Australian business optimise for first?

Google AI Overviews and AI Mode, because they sit on the searches your customers already run. Then ChatGPT, which means being visible in Bing. Perplexity is the outlier that cites top-10 pages most often.

How long before AI visibility changes?

On-site fixes such as the proof line and location details show up within weeks of a recrawl. Directory reviews, list placements and brand mentions take months to accumulate, and they are the part that moves the number most.

We run this process for every client as part of SEO with Mojo Dojo. The first 30 minutes of the work is the source map: three buying questions, every page the engine read, and which layer named your competitors. If you want to see what an AI engine reads before it answers a question about your business, that is where we start. Melbourne businesses can find us at Mojo Dojo Melbourne SEO.

Sources