The State of AI Visibility in Sri Lankan Banking

Five things the data shows

• A settled group of large banks dominates the answers. On the same 16 questions every month from December 2025 to March 2026, two banks were named or linked in about three answers in four (Section 03).
• AI leans on banks’ own websites. About 64% of citations pointed to a bank’s own site, and every month from December to March stayed between 60% and 68% (Section 04).
• Social platforms barely register. LinkedIn was about 1% of citations; Reddit was effectively 0% (Section 04).
• AI almost never cites a negative source. 0.2% of 34,532 citations carried negative sentiment (Section 04).
• AI gave no wrong email addresses in one bank’s 1,272 fact-checks. The only errors were an outdated phone number and a dead link, both gone by May (Section 05).

One bank’s question set, not a census

Every figure comes from buyer questions tracked in one bank’s SiteSignal account. It’s a close look at how AI answers a set of real banking questions in Sri Lanka, not a survey of every bank.

The question set changed over time: 3 questions in November 2025, then the same 16 from December 2025 to March 2026, then a smaller set after a platform change in May 2026. So the bank results below use December 2025 to March 2026, when the questions didn’t change. Tracking also paused for about six weeks in April 2026.

During this period our system recorded the banks an answer named or linked to, by website. An answer means one question, asked once in one engine on one day. Engines: ChatGPT, Gemini and Perplexity.

A settled group at the top

2,889 answers, 1 December 2025 to 31 March 2026. A bank’s different web addresses are merged into one (for example boc.lk and bankofceylon.com). Shares add up to more than 100% because most answers name several banks.

Below these ten, answers also named finance companies, digital lenders, foreign banks, regulators and news sites.

A sector that still runs on the bank’s own website

34,532 citations, December 2025 to March 2026, on the questions tracked.

Rates, fees and account terms mostly live in one place: the bank’s own site. So that’s what AI cites. The owned share stayed between 60% and 68% every month from December to March, with no clear trend either way. A clear, well-structured product and rates page is very likely to be exactly what a model reads.

Social sources barely register. LinkedIn was the largest social source at about 1% of citations each month. YouTube stayed under 0.3%. Reddit was effectively zero. Tone: 0.2% of citations carried negative sentiment, 44% positive and 55% neutral.

Contact details: no wrong emails, two small errors

• Whose checks these are. SiteSignal ran 1,272 fact-checks for one bank, comparing what AI engines said with the answers that bank gave us. They describe how AI treats that bank, not Sri Lankan banks in general.
• Email addresses. AI gave no wrong email addresses. When an engine didn’t know, it said “Unknown”, and every address it did give is one the bank publishes on its own website.
• The only errors. An outdated phone number and a link to a page that no longer exists, both in February and March 2026. Neither appeared from May.
• How we judge answers. “Unknown” counts as no answer. A bank’s full legal name (with “PLC”), a phone number written in a different format, or an address the bank itself publishes counts as right.

What a bank’s team can act on

• Keep product, rates and fee pages clear and current. In this sector they’re what AI cites most.
• Publish your contact details in one obvious place and keep them identical everywhere. Retire old phone numbers and fix dead links; AI keeps repeating them.
• Ask AI the questions your customers ask, in their words, and note which banks are named and which pages are cited. Then ask again next month with the same questions.

Track your own questions before borrowing these numbers. They describe one set of real banking questions in one country.