How to Find Good Backlinks for AI Search: A Data Method From 1,000 Queries

In July 2026 we pointed a scraper at Google and captured the AI Overview for 1,000 real Indonesian search queries. 930 of the 1,000 queries returned an AI-generated answer, and each answer cited 8.5 sources on average — 7,880 citation slots spread across just 1,578 domains. That last number is the one that should change how you think about link building.

If a small set of domains keeps absorbing most AI citations, then finding good backlinks is no longer about collecting authority points from any site that will have you — it is about getting your brand onto the specific pages AI engines already pull from. This post shares the full methodology we used to find those pages, including the scoring formula, so you can run the same analysis for your own niche.

Do backlinks still matter for AI search?

Backlinks still matter for AI search, but they are no longer the strongest signal — being mentioned on pages AI engines already cite matters more. Ahrefs analyzed 75,000 brands and found that branded web mentions correlate with AI Overview visibility at 0.664, roughly three times more strongly than backlinks at 0.218. A bare link does less for you than a paragraph on a trusted page that names your brand and explains what it does.

Source: Ahrefs — AI Overview brand correlation study

You can see the shift in what people search for. Queries like "AI citations vs backlinks", "how to get cited by ChatGPT", and "do backlinks still matter for AI search" have produced a wave of 2026 playbooks, and their advice converges on the same line: earn placements on sources answer engines already trust. What almost none of them tell you is how to find out which sources those are for your specific topic. That gap is what our study fills.

Which websites do AI Overviews actually cite? What 1,000 queries showed

A short list of category leaders dominates AI Overview citations: in our data, YouTube appeared in 58% of all answers, and in health, Alodokter and Halodoc were each cited in more than 81% of answers.

The setup: we generated 1,000 short, natural Indonesian queries — the kind people actually type, like "obat sariawan anak" or "hp gaming murah" — split across four categories: Health (265), Entertainment (246), Computers & Electronics (245), and News (244). For each one, we loaded live Google results in a Playwright-driven Chrome, expanded the AI Overview, opened its sources panel, and scrolled until every lazy-loaded citation card rendered. 930 queries produced an AI Overview, citing 7,880 sources in total — a median of 8 per answer.

Category	Most-cited domains (share of AI Overview answers)
Health	Alodokter (82.1%), Halodoc (81.3%), YouTube (63.4%), Hello Sehat (26.8%)
Computers & Electronics	YouTube (70.7%), Google (26.4%), Instagram (23.1%), Tokopedia (17.4%)
Entertainment	YouTube (44.0%), Instagram (30.8%), Telkomsel (16.7%), Wikipedia (15.4%)
News	YouTube (51.8%), Instagram (34.5%), Detik (14.2%), CNN Indonesia (8.6%)

Source: Intura's AI Overview citation study, July 2026 — 1,000 Indonesian queries, 7,880 citations.

Two things stand out. The ceiling is brutal: no outreach email will get you onto YouTube or into Halodoc's citation share. But the long tail is wide open — of the 1,578 cited domains, most appeared in only one or two answers. Between the untouchable giants and the one-off citations sits a middle tier of domains that get cited consistently and still accept contributions, quotes, and partnerships. Those are your targets, and the rest of this post is how to find them systematically.

What makes a backlink "good" for AI search?

A good backlink for AI search comes from a domain that already gets cited for queries close to yours, at an authority tier you can realistically land. Domain Authority alone tells you none of that. Four signals do:

• Citation rate — the share of AI answers in your category that cite the domain. This is direct evidence the domain sits in the engine's trust set, not a proxy for it.
• Semantic fit — whether the queries the domain gets cited for look like your queries. A health portal cited 200 times does nothing for a gaming-phone brand.
• Citation position — how early the domain appears in the source list. Earlier positions suggest the engine treats it as a primary source rather than supporting material.
• Reachability — whether the site would ever actually link to you. YouTube's 58% citation rate is useless in your outreach spreadsheet; a niche review site's 12% is gold.

How do you score backlink targets for AI citations? The full methodology

The method has five steps: build a realistic query set, capture live AI answers, compute per-domain citation rates, profile each domain semantically, and score every candidate with a weighted formula. Here is each step as we ran it.

Step 1 — Build a query set that mirrors real searches

We used Gemini 2.5 Flash with Google Search grounding to generate 1,000 unique Indonesian queries, steered with few-shot examples of how people actually type: short, informal, 2–6 words, usually no question mark ("gejala tipes", "rekomendasi film netflix"). The generator round-robins across the four categories and holds roughly 75% of queries to informational intent. Whatever your market, the principle is the same: calibrate the query set to real typing behavior, not to marketing-speak keywords — polished queries produce a polished, unrepresentative citation map.

Step 2 — Capture the AI answers, including every source

We drove real Chrome with Playwright: type the query into Google's search box, wait for the AI Overview, expand it, open the "show all" sources dialog, and scroll it until every lazy-loaded citation card renders. That last part matters — the collapsed view shows only a handful of sources, and stopping there undercounts citations badly. Each query saves the title, source, snippet, URL, and domain of every citation. 93% of our queries returned an AI Overview.

Step 3 — Compute citation rate per domain, per category

Count each domain once per query it appears in, then divide by the number of queries in that category. Alodokter cited in 211 of 257 health answers is an 82.1% citation rate. Do this per category, not globally — an engine's trust in a domain is topical, and mixing categories hides that.

Step 4 — Give every domain a semantic profile

Concatenate all the queries a domain was cited for into one document, then embed it with a multilingual embedding model — we used BAAI's open-source BGE-M3 with normalized vectors. Every domain becomes a point in semantic space, positioned by what the AI engine trusts it for. Two tech sites can have identical citation rates and completely different profiles: one cited for budget phones, the other for enterprise laptops.

Step 5 — Score every candidate with a weighted formula

Embed your target query with the same model, compare it against every domain in your category, and rank with three weighted components:

One filter before scoring: drop every domain above the 95th percentile of citation rate in its category. Those are the YouTubes and the Halodocs — domains that dominate answers but will never link to you. Removing them keeps the ranked list actionable instead of aspirational. For a query like "hp gaming murah" (cheap gaming phone), the output is a top-20 list of mid-tier, topically matched domains, each already proven to appear in AI answers about your subject.

How do you turn a ranked domain list into actual citations?

Pitch mentions and placements, not bare links. The ranked list tells you where outreach effort pays off; these four plays convert it:

• Contribute expertise to mid-tier cited domains — guest articles, expert quotes, and data contributions on sites in your top 20. A named brand mention with context beats an anchor-text link, per the Ahrefs correlation above.
• Get into roundups that already get cited — a "best X" listicle that appears in AI answers passes visibility every time the engine reuses it. One placement, recurring citations.
• Build presence on the platforms that dominate — YouTube appeared in 58% of our answers and led three of the four categories. A channel that answers your category's questions is a citation asset, not a nice-to-have.
• Keep cited pages fresh — SE Ranking's analysis found pages updated within the last two months earn 5.0 citations on average, versus 3.9 for pages older than two years.

Source: Superlines — AI search statistics 2026

Where should you start this week?

Start by measuring which domains AI engines already cite in your category — even 100 queries checked by hand will show you the pattern — then aim outreach at the middle tier the formula surfaces. The order of operations is query set first, citation data second, outreach last. Most teams do it backwards: they pitch domains chosen by Domain Authority and hope AI engines agree.

Running this continuously across ChatGPT, Gemini, Perplexity, and Google AI Overviews is the unglamorous part — the citation map shifts as models update, so a one-off scrape goes stale. That is the job Intura's AI Brand Mention Tracking automates: it monitors which sources AI engines cite in your category and how your brand shows up in them, so your backlink targets come from this month's data, not last quarter's.