Low-Budget Research AI Can Cite: Lessons From n=127

By Tessar Napitupulu, founder and CEO of Arfadia, the agency that introduced GEO in Indonesia, and a member of the Forbes Agency Council. October 2026.

Low-budget research gets cited by AI when a third party can check it: a stated sample size, a stated method, a permanent identifier, and a citation that takes one click to copy. Our State of SEO Indonesia 2026 had all four, built on 127 survey respondents and a free DOI from Zenodo. Since March 2026, about 15 leads a day have reached us, and 83% of them say an AI assistant recommended us. I can't credit all of that to one report. I'll explain exactly why below, and what we did that cost almost nothing.

What actually happened after we published

Two things are easy to point at.

First, coverage. Tempo wrote about our three research reports, each registered with a DOI on Zenodo, and the State of SEO Indonesia 2026 is one of them. The other two are the AI Citation Rate Report 2026 and the Digital Marketing Benchmark Indonesia 2026. Getting covered because of research feels very different from getting covered because you're selling something. Journalists treat it differently, too.

Second, leads. Since March 2026 we average about 15 inbound leads a day. We ask every one of them the same question, where did you hear about us, and 83% answer with an AI recommendation. The assistants they name are all over the place: ChatGPT, Gemini, Meta AI, Copilot, Claude, Perplexity and DeepSeek.

Now the honest part. That 83% is self-reported, from a single question, and the report is only one of the things AI systems read about us. They also read our client case studies, media coverage and agency directory profiles. So no, one report did not produce 15 leads a day. What I can say with a straight face is narrower: the report gave AI something a third party can verify. It has a DOI. It has real numbers. It has a written method. Agencies get named in AI answers when there's a checkable, named thing attached to them, and research is one of the cheapest named things you can build.

The report also keeps paying rent internally. We use its numbers in client proposals, in my Forbes Agency Council writing and in social content. One finding gets quoted more than any other: only 23% of the Indonesian businesses we surveyed have a formal Generative Engine Optimization strategy, which means 77% don't. Specific, sourced, easy to repeat. That's the kind of sentence an AI answer can lift.

The 10 low-budget tactics we actually used

None of this needed a big budget. Most of it needed discipline, and a willingness to publish numbers that weren't as round as we wanted them to be. Here's the full list, in roughly the order we did them.

1. Keep the questionnaire short enough for a phone

The survey had 42 questions split into 13 short sections, from respondent screening through budgets, KPIs, technical SEO, AI use and outlook. We designed it mobile-first, because most of the people we wanted (business owners and marketing leads) answer surveys on their phones, between meetings. A long desktop form loses exactly those people.

2. Recruit through a panel and through people who already know you

We ran fieldwork in January and February 2026 through an online panel plus outreach to our warm network. The warm network definately matters more than people think. Someone who has met you is far more likely to finish question 42 than a stranger is.

3. Set the minimum sample before fieldwork, then publish the real n

Our target was 150. Our floor, decided before the first response came in, was 120. We closed at 127, below target but above the floor, so we published it and wrote n=127. Not "100+", not "nearly 150". Here's why that matters for AI visibility: a precise, stated sample is a checkable fact. A rounded-up one is a liability the first time someone reads the methodology.

4. Use quotas so one segment can't drown out the rest

We used disproportionate stratified sampling, which is a fancy way of saying we set quotas by company size on purpose. The targets were 40 enterprise, 55 mid-market and 55 small businesses. We got 31, 52 and 44. The report says so, in a table, and it says the design was built for comparison between segments, not for estimating the whole Indonesian business population.

5. Borrow context from free public data

Survey answers tell you what respondents think. Context tells you whether it matters. We pulled more than 40 secondary sources, many of them free and public, like APJII internet penetration data, BPS statistics and Google's own published figures. Every third-party number links back to its original source, not to a blog roundup that quoted it.

6. Register a free DOI

Zenodo, the open research repository operated by CERN, issues a DOI at no cost. A DOI does not mean peer review, and we never say it does. What it does is make the report permanently citable, and Zenodo's versioning means every revision gets its own record. You can update the numbers. You just can't update them quietly.

7. Don't put the report behind a form

The report is readable without an email address. Gating research feels smart for lead capture, but a crawler can't fill in your form and neither can a journalist on deadline. (Our ebooks are gated. Our research isn't. Different jobs.)

8. Make citing it a one-click job

The report page has a cite button that copies the reference in APA or BibTeX. It took very little time to build. It removes the one step where most people give up and paraphrase you without credit.

9. Publish in more than one language

The report page exists in English, Indonesian and Chinese. AI assistants answer in the language of the question, and a page that only exists in English is easy to skip when someone asks in Bahasa Indonesia.

10. Cut one report into many pieces

One report became an 8-slide LinkedIn carousel, short videos, and posts on every channel we run. Each piece points back to the same DOI. That repetition across platforms is boring to plan and it works anyway.

Put side by side, here's what each tactic gives the people (and systems) who might check your work:

Tactic	What we did	Signal it creates	Who can check it
Short, mobile-first survey	42 questions, 13 sections	Completed responses from busy decision makers	Anyone reading the method section
Panel plus warm network	January to February 2026 fieldwork	A stated recruitment method	Journalists, researchers
Floor set before fieldwork	Target 150, floor 120, result 127	An exact n instead of a rounded claim	AI systems comparing sources
Size quotas	31 enterprise, 52 mid-market, 44 small	Segment comparisons that hold up	Analysts, competitors
Free public context	40+ sources, linked to originals	Claims traceable past your own page	Editors, fact checkers
Free DOI	Zenodo registration, versioned	A permanent, named identifier	Anyone with doi.org
No form	Ungated report pages	Content crawlers can actually read	AI crawlers, journalists
Cite button	APA and BibTeX in one click	Credit that arrives with the right name	Writers, students, researchers
Three languages	English, Indonesian, Chinese pages	Presence in non-English questions	AI assistants answering locally
Repurposing	Carousel, short videos, posts	The same DOI repeated across platforms	Everyone who scrolls past it

If I had to keep only two, I'd keep the DOI and the cite button. They cost almost nothing, and they change how people treat an agency report. It stops reading like a brochure and starts reading like a reference.

The AI-sourced deal we still lost

This is the part I most want other founders to hear, because it's the part nobody posts about.

A large company messaged us on WhatsApp. They needed a social media vendor. When we asked how they found us, they told us straight away: ChatGPT and Claude. So far, a perfect AI visibility story.

Then it became a formal tender. Briefing session, full proposal with a hard copy, clarification meeting, price negotiation. We asked about budget in the very first chat. They couldn't share it, because their procurement rules didn't allow it.

We lost. When we asked why, they told us our technical score was good, but our price was above their limit.

So here's the lesson, and it's a simple one. AI can bring a buyer to your door. It doesn't know their budget. An AI recommendation filters for credibility, not for price fit, which means a lead from ChatGPT still has to be qualified exactly like a lead from any other channel. Sometimes harder, honestly, because the buyer arrives already convinced you're credible and you can forget to check everything else.

The good news came later. They invited us back for a new procurement. Showing up in the AI answer got us on their vendor list. Whether we win the next one is a pricing conversation, not a visibility one.

How we measure AI visibility that actually matters

The most honest metric is also the simplest one. Ask every lead where they heard about you, and write the answer down. That's where our 83% comes from. No monitoring tool can give you that number, because no monitoring tool sits in your WhatsApp inbox.

On the monitoring side, these are the rules we hold ourselves to:

• Lock the prompt panel on day one. We agree the list of prompts at the start of a contract and record the baseline, even when it's zero. If prompts can be swapped halfway through, the numbers are way too easy to make look good.
• Separate branded prompts from category prompts. Our own data taught us this one. When we analyzed 7,771 AI citations in our Promptwatch account in September 2026, our visibility score sat between 79 and 86 on prompts that already contained the name Arfadia, and at zero on nine "how do I" category prompts. Looking strong for people who already know your name doesn't bring you new buyers.
• Measure share of voice against competitors who really show up. The default competitor set in our tool was a group of American agencies that never appeared in answers for Indonesian buyers. Beating them meant nothing. We replaced them with the names that actually appear in local answers.
• Track the share of non-branded traffic. For one manufacturing client, 81% of organic traffic is non-branded. Those visitors didn't know the brand before they arrived, which is the whole point.
• Don't report a number that isn't live yet. Our AI platform traffic metric showed sample data until the integration was connected. We left it out of reports until it was real.

All of this sits inside RoGEO, the framework we developed to measure Generative Engine Optimization: citation frequency, reference depth and revenue attribution. The first two are about whether AI mentions you and how seriously. The third is the one that matters to a CFO, and it's also the one that brings us back to that first, boring question. Where did you hear about us? If you want to see how we run this for clients, our GEO service page lays out the process.

When this approach won't work

Look, this isn't a magic formula. A few situations where it falls short:

• You need to prove one asset caused one result. You can't, at least not cleanly. Leads come from a mix of signals, and self-reported answers are honest but imprecise.
• You want population-level claims. A sample of 127 with deliberate quotas is built for comparing segments. It doesn't describe every business in a country, and you shouldn't write as if it does.
• You need results next month. AI systems pick things up slowly. In our experience the first signals show in Perplexity after 4 to 8 weeks, measurable movement starts around month three, and consistent citation frequency takes 4 to 8 months.
• Your sales process can't qualify budget. Visibility fills the top of the funnel. If nobody asks the budget question early, you'll write proposals for buyers you can't win, like we did.

Where to go deeper

I wrote about measurement, the RoGEO framework and what AI systems need before they'll cite a brand in my book Cited or Silent: The Definitive GEO, AEO & AI Visibility Playbook. The free digital edition is available here, and the paperback is on Amazon, with ebook editions on Google Play and Apple Books. And if you just want the data behind this article, the State of SEO Indonesia 2026 is open, no form required.

---

Frequently Asked Questions

Can a small survey really get cited by AI assistants?

Yes, if it is verifiable. AI systems and journalists look for a stated sample, a stated method and a stable place to cite. A survey of 127 respondents with a DOI and a written methodology gives them all three. Size matters less than whether someone else can check what you did.

What sample size do you need for a citable industry survey?

There is no universal number. What matters is setting a minimum before fieldwork and publishing the exact result. We set a floor of 120 and a target of 150, closed at 127, and reported n=127 together with the quota table by company size.

Does a Zenodo DOI mean the research is peer-reviewed?

No. A DOI makes a publication permanently citable and Zenodo keeps a public record of each version, but it is not peer review. Our reports are open access on Zenodo with registered DOIs, and we do not describe them as peer-reviewed.

Should original research be gated behind a form?

If you want it cited, no. Crawlers and journalists cannot fill in a form. We keep our three research reports ungated and use our ebooks as the only lead magnet that asks for contact details.

How do you know a lead came from an AI recommendation?

We ask every inbound lead where they heard about us. Since March 2026, 83% of roughly 15 leads a day have named an AI assistant. It is self-reported data, so we treat it as a strong signal rather than exact attribution.

Why separate branded and category prompts when tracking AI visibility?

Because they measure different things. In our own September 2026 data, visibility was 79 to 86 on prompts that named Arfadia and zero on nine category prompts. Branded prompts show recognition among people who already know you. Category prompts show whether new buyers find you.

How long before AI assistants start citing new research?

In our experience, early signals appear in Perplexity after 4 to 8 weeks, measurable movement starts around the third month, and consistent citation frequency takes 4 to 8 months.

Does an AI recommendation mean a lead is qualified?

No. An AI recommendation filters for credibility, not budget. We lost a formal tender to a company that found us through ChatGPT and Claude because our price was above their limit, even though our technical score was good.

---