Which home loan lenders does AI recommend in Australia?

When you Google a home loan in Australia, the big four banks take the page. A mid-tier lender spending on awareness — Macquarie, ING, Athena, Unloan, UBank, Reduce — can buy recognition and still not appear when a shopper asks ChatGPT or Claude for a recommendation. The brands investing in being known are not always the brands an AI names. The question is which ones it does name, and why. A lender counts as recommended on a call when the engine names it in its answer; its rate is the share of calls that name it.

At the top, AI echoes market position

Commonwealth Bank leads the eight tracked lenders on both engines — 42.2% of calls on ChatGPT and 64.5% on Claude — and it leads the consumer funnel too: a 2025 Conjointly survey puts Commonwealth Bank at 17.9% preference share, more than three times the nearest rival, with 52.5% awareness and 32.3% consideration.^[1] Westpac, the second major bank, sits at 37.0% on ChatGPT and 36.8% on Claude. For the two big banks, AI recommendation is directionally consistent with the Conjointly survey — CommBank leads both.

Below CommBank, the shortlist follows a Canstar award

The assumption that AI echoes market share breaks at the next tier down. Of the six mid-tier lenders we tracked, the two holding a current Canstar Outstanding Value or Bank of the Year award — Macquarie and Unloan — averaged a 28.7% recommendation rate across both engines. The four without one — ING, Athena, UBank, Reduce — averaged 12.1%. That is a 2.4× gap. Macquarie sits in the band on both engines (35.0% ChatGPT, 36.0% Claude). Unloan sits in the band on Claude at 34.1% (9.6% on ChatGPT, below the band). Macquarie holds the 2025 Bank of the Year for fixed-rate home loans^[2]; Unloan holds the 2026 Canstar Outstanding Value Award — Variable Home Lender.^[3] The mid-tier brands with the highest recommendation rates are not random — they are the ones a comparison site has already singled out.

An award is a page the AI already has to retrieve

A Canstar award is not a signal the model reads off a trophy. It is a generator of pages. An award produces editorial explainers, ranked comparison tables, and linked lender profiles — the kind of documents a retrieval-based answer engine reads. The three AU comparison sites tell different stories. InfoChoice's home-loans page lists all eight tracked lenders, so it does not separate the ones an AI names from the rest.^[4] RateCity now redirects to canstar.com.au — its standalone leaderboards are gone.^[5] Canstar is the column that discriminates: of the mid-tier, only Macquarie and Unloan currently hold an award.^[3] In this data, the lenders an AI names are the ones a comparison site has already written up. We measured which lenders each engine named, not which pages it fetched, so whether those pages cause the naming is a question for a retrieval-logging study, not this one.

Claude named seven of eight; ChatGPT named three

The two engines agree on CommBank, Westpac, and Macquarie — all three land in the band on both. After that, the lists diverge. Claude placed 7 of 8 real lenders in the band; ChatGPT placed 3. Unloan, Athena, Reduce, and ING all appeared on Claude’s shortlist but fell below the 15% floor on ChatGPT. Both fictional controls — Kestrel Home Loans and Veldra Finance — scored 0% on both engines, confirming neither engine hallucinated a lender.

Eight questions, 816 calls, one number locked before we looked

We asked both engines the same 8 questions a home-loan shopper would type. Each question ran across 8 real lenders (the two big-bank majors plus six mid-tier) and 2 fictional control lenders we invented as a quality check, 17 repeats each, on both ChatGPT and Claude with live web search, on day-0 only. That is 408 calls per lender per engine, 816 calls in total. A lender is in the band when its recommendation rate sits between 0.15 and 0.85. Before we read a single result we locked one go/no-go number: at least three mid-tier lenders in the band on one engine, or the study stops at a one-batch observation. Five of the six mid-tier lenders cleared that bar on at least one engine. Two registered checks confirmed the formula’s intervals behave as advertised (98.5% coverage, above the 90% threshold) and found no evidence that recommendation rates scatter more than chance across day-0 batches (observed 90.78, under the 117.24 chance-only high mark).

• "best home loans Australia"
• "top home loan lenders in Australia"
• "best variable rate home loan Australia"
• "lowest interest rate home loan Australia"
• "best home loan for first home buyers Australia"
• "best fixed rate home loan Australia"
• "recommend 5 home loan providers in Australia and say why"
• "best offset mortgage Australia"

Two weeks later, the shortlist held

We asked the same eight questions again two weeks later, on day 14 (the one-week check was missed — see the limitations below). Of the 44 lender-question pairs that sat in the band on day 0, 43 moved no further than a same-day repeat already moves on its own. One moved more: Commonwealth Bank, on one of the eight questions, jumped 33 percentage points on Claude. At this sample size the check only reliably catches a swing bigger than about 30 percentage points, so it rules out a large shift, not a small one — and one exceedance out of 44 checks is about what you'd expect from chance alone (roughly 2), not a sign the model changed course.

Two dates, six mid-tier brands, one country

The 2.4× Canstar gap is a descriptive on six mid-tier lenders, not a registered bet and not a cause — having an award and being recommended move together in this data, but six brands cannot separate correlation from coincidence. This is one country, two engines, and — with the two-week check above — two dates, not a continuous track; the one-week check was skipped, disclosed above. Eight brands is not the Australian market; lenders we did not track, including ANZ and NAB, are not scored. These queries ran against the OpenAI and Anthropic APIs with live web search, not the consumer ChatGPT or Claude products (the consumer-vs-API gap is detailed in the limitations below).

The pattern points to the comparison site, not the lender’s own site

A mid-tier lender without a comparison-site award has fewer stand-out pages for an AI to cite, and in this data it shows up at less than half the rate of the two that hold one. The shortlist an AI names follows the comparison-site layer that already existed: the lenders that comparison editors had already singled out are the lenders with the pages an AI retrieves, and the lenders with pages are the lenders it names. For a mid-tier brand, the pattern in this data is not about its own site; it tracks whether a comparison site has already written it up.

How we ran this

What we measured: 8 Australian-specific home-loan questions × 8 real lenders (2 major-bank majors + 6 mid-tier) + 2 fictional control lenders, 17 repeats each, on ChatGPT and Claude with live web search, day-0 only. A lender is in-band when its recommendation rate is between 0.15 and 0.85. The registered metric is per-lender recommendation rate pooled across the 8 questions, with Wilson 95% intervals labelled illustrative. Because one call can name several lenders, outcomes within a call are not independent, so these intervals are descriptive, not inferential. Australian home loans was the first category through the stability study's pilot gate — a high-consideration market with a defined mid-tier — so the choice was set before any rates were read.

When: Data collected 21–22 July 2026 (day-0 window) and 3 August 2026 (day-14 window); the one-week window was missed. Day-0 analysis published 22 July 2026; the two-week stability check added 3 August 2026.

References