How to rank in AI search

Let the right bots in

If an AI engine can’t reach your page, it can’t cite it, and every other step in this guide is wasted effort. In our study, ChatGPT ran searches like site:aia.co.nz New Zealand life insurance medical examination underwriting and site:partnerslife.co.nz financial strength rating AM Best 2026 Partners Life in the middle of writing its answers. If either site had been blocking OpenAI’s search crawler, those searches would have come back empty and someone else’s page would have filled the gap.

Check these five things:

1. Search bots vs training bots. Most AI companies run separate crawlers for search and for training their models. OpenAI, for example, uses OAI-SearchBot for ChatGPT search and GPTBot for training (OpenAI). You can opt out of training and still show up in search, and that’s what most publishers now do: across Cloudflare’s network, training bots are blocked far more often than answer bots (TechnologyChecker, 2026).
2. Your CDN and firewall, not just robots.txt. Plenty of sites say “come on in” in robots.txt, then block AI bots at the firewall. That’s leaving the front door open and padlocking the gate.
3. Content that loads without JavaScript. Vercel’s analysis found that none of the major AI crawlers rendered JavaScript (Vercel). If your prices, policy details or FAQs only appear after scripts run, AI may never see them. Try viewing your key pages with JavaScript switched off; what’s left is roughly what the bots get.
4. Bing Webmaster Tools and IndexNow. Bing’s index powers Microsoft Copilot and feeds other AI search tools, so verify your site there and use IndexNow to push updates quickly.
5. Your server logs. They show which AI bots visit, how often and what they fetch. It’s the closest you’ll get to seeing your site through the robots’ eyes.

Crawlability also feeds the second route into AI answers: training data. A site that’s hard to crawl today is less likely to shape what the next generation of models learns about your brand.

Keep winning at classic SEO

AI engines don’t invent their sources. They search, and what they find depends on search indexes. Ahrefs found that 76% of the pages cited in Google’s AI Overviews also rank in the top 10 organic results (Ahrefs), and one study found ChatGPT searched the web on 31% of prompts (Nectiv via Search Engine Land). Strong SEO is still your ticket in.

But the engines don’t all read the same shortlist. In our study, Google’s AI Overviews cited 361 different pages, while ChatGPT and Gemini cited about 170 each. ChatGPT also had its own taste: a third of the pages it cited belonged to insurers, compared with about one in seven for AI Overviews and fewer than one in ten for Gemini. It leaned harder on regulators and public guides as well.

Shares are of the different pages each engine cited. Source types are our own grouping of the cited domains. Highlighted cells show the highest share in each row.

The practical upshot: keep doing good SEO for Google, make sure you’re properly indexed in Bing, and track each engine separately rather than assuming a win in one is a win in all three.

Map prompts by intent, not keywords

This is where our data gets interesting. The same brand can be a star on one kind of question and nearly invisible on another.

• AIA was visible in 87.5% of comparison answers but only 16.7% of cost answers.
• Asteron Life went from 60.3% on comparison questions to 8.9% on informational and 4.2% on cost.
• MAS barely featured in comparisons (5.0%) but came top on cost questions (31.3%).
• Informational questions had no clear winner at all. Nobody passed 23%, and a comparison site (Quotehub) led the pack.

Insurer Comparison or adviser site. Cost questions are a small group (8 questions), so treat them as indicative.

If you only tracked “best life insurance NZ”, you’d think AIA owns the category. If you only tracked cost questions, you’d put your money on MAS. Both are true, for different customers at different moments. That’s why AI search research starts with a prompt library rather than a keyword list.

Build a prompt library

A good prompt library covers the questions your customers ask at every stage, in the words they use.

• By persona: first-home buyer, young family, self-employed, nearing retirement.
• By stage: “Do I need it?”, “How does it work?”, “Who’s best?”, “How much?”.
• By intent: comparison, informational, cost and branded questions (“Is Partners Life better than AIA?”).
• In your customers’ own words: Kiwis talk about KiwiSaver, ACC and advisers. Some add “NZ” or their city and plenty don’t, so track questions exactly as people ask them.

Mine your Search Console queries, sales and support calls, “People also ask” boxes and, where you can see them, AI fan-out searches. Between 50 and 150 prompts is a sensible start for most brands.

Build one clear answer page per big question

Remember MAS, the cost-question leader from step 3? It’s the underdog story of our study, and much of the credit goes to one page: “How much does life insurance cost?” on the MAS website was cited for seven of our questions, on all three engines. One clear answer to one big question beat brand size.

The same pattern runs through the most-cited pages. Moneyhub’s life insurance guide was cited for 31 of our 72 questions, 98 times in total. AIA’s product pages did well too, backed by the most visible brand in the study. But most of the other insurer pages that earned citations answered one specific question: Partners Life on buying direct vs through an adviser, Chubb on life insurance for single people, MAS on whether you need life insurance at all. For smaller brands especially, a page that answers one question well was the most reliable way in.

What a citable answer page looks like

• Answer first. Put the direct answer in the opening sentences, then add depth. One analysis of 18,012 ChatGPT citations found 44% came from the first 30% of a page (Search Engine Land).
• Be specific. Use real numbers, dates, examples and conditions. Dates and numbers were among the strongest predictors of a ChatGPT citation in that same analysis.
• Show your working. Say where your figures come from and when you last checked them. AI engines, like people, trust a claim they can verify.
• Use tables for comparisons. Side-by-side options are easy for an engine to lift and easy for a reader to scan.
• Publish something only you know. When AI described Life Covered, it singled out its “specific, regularly updated quote examples”. Original data, first-hand experience and expert opinion are what make your page worth citing over ten others that say the same thing.

Get onto the pages AI already trusts

Here’s a humbling number for anyone who works on a brand website. In our study, insurers’ own sites earned just 16% of the citations. The other 84% went to pages the insurers don’t own.

Source types are our own grouping of the 222 cited domains.

Comparison, broker and adviser sites took the biggest share (43%). The ten most-cited domains captured 42% of all citations between them, led by Moneyhub, Quotehub and Life Covered. Regulators and public guides such as the FMA, Consumer NZ and Sorted earned fewer links, but ChatGPT went looking for them by name. McKinsey found something similar on a much bigger scale: a brand’s own sites make up only 5 to 10% of the sources AI search references (McKinsey).

In other words, AI trusts you more when someone else says it. The job is to make sure the pages it already trusts say accurate, positive things about you.

Run a citation gap analysis

1. List the pages AI cites for your priority prompts. Your tracking will give you this.
2. Mark the ones that mention competitors but not you. These are your gaps.
3. Work out why. An outdated comparison? A missing listing? A writer who’s never heard of you?
4. Fix what you can. Update listings, offer data or expert comment, pitch reviews and correct errors.

Let your own data guide your channels, not generic advice. “Get on Reddit” appears in almost every AI search guide, yet in our latest snapshot Reddit earned about 1% of citations for NZ life insurance. For now, being covered accurately by the handful of comparison sites AI keeps coming back to matters far more in this category.

But treat that as a snapshot, not a rule. The sources AI leans on to judge consensus aren’t fixed. If a busy, well-informed Reddit thread about the best life insurers in New Zealand took off tomorrow, it could start ranking in Google and Bing and soon become a regular AI citation. That’s why tracking which pages get cited, over time, matters as much as tracking your own visibility.

Beware the listicle trap

Third-party “best of” pages get cited a lot. Life Covered’s “best life insurance companies” page alone was cited for 19 of our questions. It’s tempting to publish your own “best life insurers in NZ” list with yourself at number one. Resist the urge. As we saw earlier, your own list can get cited as a source while you’re left out of the recommendation, and Lily Ray also found that Google has demoted sites leaning heavily on self-promotional lists. You end up doing your competitors’ PR for them. Earn a spot on other people’s lists instead.

A word on cause and effect: Ahrefs’ study of 75,000 brands found branded web mentions correlate more strongly with AI Overview visibility (0.664) than referring domains do (0.295). But correlation isn’t causation. Mentions and visibility both tend to follow from being a well-known, well-regarded business. Chase mentions by being worth mentioning, not by manufacturing them.

Check how AI describes you

Being mentioned is only half the story. What does AI actually say about you? We fed every brand mention from our study back through OpenAI and asked it to summarise how each brand was positioned and what were the key attributes attributed to each brand in these mentions. The results read like brand briefs written by someone who has read the whole internet.

What AI says about each brand, based on every brand mention. Summaries were generated with OpenAI and condensed by us; quoted attributes are the AI’s own words. They reflect what AI is saying, and may include out-of-date claims. Five tracked brands (Smart Insure, InsureNZ, Global Finance, Momentum Life and Foundation Life) weren’t mentioned enough to summarise.

Look at how each brand’s AI-described strength lines up with where it wins. Pinnacle Life is described as the low-premium, direct option, and it was the second most visible brand on cost questions. Partners Life, AIA and Fidelity Life are described in terms of claims records, breadth and scale, and they dominate comparison questions. MAS is described as a mutual with high customer satisfaction and “competitive premiums in some cases”, and it topped the cost questions. Your AI positioning isn’t just a vibe. It decides which questions you win.

Notice, too, how specific the AI gets. It quotes claims acceptance rates, financial strength ratings, policy features and pricing details, pulled from pages that may or may not be current. If any of that is wrong or out of date, it’s being repeated to your customers every day.

How to audit and fix your AI positioning

1. Ask the questions your customers ask and read what comes back about you, across several engines and several runs.
2. List every factual claim: prices, ratings, features, claims statistics, ownership and product names.
3. Trace wrong or outdated claims to their source. The problem is usually on someone else’s site.
4. Correct the source. Update your listings, ask publishers to refresh old comparisons, and publish the correct facts clearly on your own site.
5. Keep your core facts consistent everywhere, so there’s one version of the truth for AI to find.

Measure honestly

If someone shows you a screenshot of their brand at “number one in ChatGPT”, smile politely. SparkToro’s research found there’s less than a 1 in 100 chance that AI tools give the same list of brands twice for the same prompt, and less than a 1 in 1,000 chance of the same list in the same order (SparkToro, 2026). One screenshot tells you about one person, at one moment.

What you can measure is how often you appear across a large set of questions, asked repeatedly. That’s the approach we used, and it works as a template:

• A fixed prompt library, split by intent and persona.
• Every major engine your customers use, tracked separately.
• Clean, logged-out sessions in each engine’s web interface, so personal context doesn’t skew the results.
• Repeated runs over time, so one-off wobbles average out.
• Mentions and citations reported separately, plus accuracy and sentiment.
• Every cited page logged, so you can see who’s shaping the answers.

That clean-session approach gives you a baseline, not the exact answer every customer sees. Logged-in users bring their chat history, memories and location with them, so their answers will vary around that baseline. What the baseline tells you is whether the underlying consensus about your brand is moving in the right direction.

Brand rankings in our study have been very consistent, and the most influential pages were rock solid: every one of the 50 most-cited pages was cited again each time we checked. The churn happened in the long tail, where a quarter of cited pages were new and 179 pages seen earlier had dropped out.

The pages that shaped the AI answers in our study are relatively few and stable. That may vary from industry to industry, depending on product cycles and how active third-party publishers are, and as step 5 showed, new sources can break in at any time. Keep tracking where AI draws its conclusions from so you can adjust your strategy accordingly.

Tools worth adding

• Google Search Console now has a generative AI performance report showing impressions for your pages in AI Overviews and AI Mode, by page, country and device. It’s rolling out gradually, so not every property has it yet (Google).
• GA4 now includes an AI Assistant channel in its default channel grouping, so you can connect AI referrals to leads and revenue (Semrush).
• Realistic timelines: technical fixes can show up within weeks once crawlers revisit. Content and third-party coverage usually take a few months to shift the consensus.