Most case studies about AI search visibility start with a brand that already has an audience. This one does not.
PT Garuda Yamato Steel is a joint venture between Yamato Kogyo, Siam Yamato Steel, PT Hanwa Indonesia, and PT Gunung Raja Paksi. It runs a plant with one million tonnes of annual capacity and produces Indonesia's first seismic grade structural steel. Serious operation, serious product, serious partners.
And in December 2024, as far as any search engine or AI assistant was concerned, it did not exist.
Not underperforming. Not ranking badly. Just absent. Fourteen referring domains. A thousand organic visitors a month. Zero AI Overview placements. When a procurement manager asked ChatGPT which steel manufacturer to use in Indonesia, this company was never named, because from the model's point of view there was nothing to name.
Twenty months later the same domain appears in 680 AI-generated responses across six generative platforms, and holds the number one AI Overview position for 97 keywords.
Here is how that happened, with every figure sourced from tools anyone can check. The complete case study carries the full data set.
The Starting Point: What the Audit Actually Found
Arfadia published a complete audit before touching anything. That matters more than it sounds, because it means every number below is a measurement against a documented baseline rather than a claim made after the fact.
Baseline, 3 December 2024
| Metric | Value at baseline |
|---|---|
| Referring domains | 14 |
| Monthly organic traffic | 1,020 |
| Keywords in Google top 3 | 16 |
| Monthly search impressions | 13,100 |
| Quality backlinks (Moz) | 0 |
| Total social signals | 0 |
| AI Overview placements | 0 |
What That Meant in Practice
A third party site audit returned an overall grade of D, with Links and Social both graded F. Moz Domain Authority sat at 2.
The company existed physically. It did not exist as a machine readable entity. Search engines had almost no external signal that it was real, and generative AI systems had nothing at all to retrieve.
This is the part people underestimate. A new company is not competing badly in AI answers. It is not in the consideration set at all, and there is no feedback to tell you that.
Why This Is Harder Than Scaling an Existing Brand
Most published AI visibility case studies take a brand that already has media history, an existing backlink profile, and some presence in AI answers, then improve it. Moving an established company from 712 to 3,235 AI citations is a multiple, and multiples are easier than they look.
Garuda Yamato Steel had no brand entity, no media history, no backlink profile, and no AI presence whatsoever. Moving a company from zero is not a multiple. It is a category change.
The reason is mechanical. Generative engines resolve whether a company is real before they will recommend it, and that resolution comes from third party corroboration rather than from the company's own website. You cannot publish your way out of it from your own domain.
The Method: Five Workstreams, Not a Content Calendar
This result traces to a repeatable method. Each workstream below addresses a specific mechanism in how generative engines decide which companies to name.
Workstream 1: Entity Foundation
A generative engine cannot recommend a company it cannot resolve. Consistent name, address, product, certification, and ownership signals were established across the website, structured data, institutional registries, and third party databases before a single article was published.
This is unglamorous work that produces nothing visible for months. Skip it and everything after it fails to stick.
Workstream 2: Technical Remediation for AI Crawlers
Mobile performance was rebuilt from 38 to 92, desktop from 64 to 99, and Best Practices moved from 75 to 100. No AI crawler blocked, no snippet blocking directive applied, and the server returns a complete response to GPTBot so crawlers read full pages without executing JavaScript.
Indexed pages grew from 164 to 403 over the engagement.
Workstream 3: Specification Grade Technical Content
130 bilingual articles, each reviewed by Garuda Yamato Steel engineers before publication. The editorial standard was deliberately narrow: answer the question a structural engineer actually types while sizing a member for a live project.
Dimension tables, weight charts, and tolerance figures are factual and unambiguous, which is exactly what a generative engine can lift into an answer. Marketing prose is almost never cited, because there is nothing in it to extract.
The content ranks because it is technically correct, and it is technically correct because engineers checked it.
Workstream 4: Comparison Format Earned Media
This one surprises most clients. The most effective media placement was never the praise piece.
It was the article benchmarking Garuda Yamato Steel against named peers, because that is precisely the structure a language model extracts when a user asks it to compare suppliers. Three of the AI queries where the brand now ranks first trace directly to this content family.
Workstream 5: Continuous Multi Engine Measurement
7,303 dated prompt tests across multiple devices and engines, with screenshot capture showing visible system dates. Positions are re-verified monthly against third party exports rather than asserted. Where a number cannot be evidenced, it is not reported.
GEO Results: From Zero to 234 Owned AI Overview Placements
This is the centre of the case.
AI Overview Ownership Over Time
| Date | AI Overview placements owned |
|---|---|
| 3 December 2024 | 0 |
| 1 June 2025 | 53 |
| 31 December 2025 | 104 |
| 31 March 2026 | 130 |
| 30 June 2026 | 203 |
| 21 July 2026 | 234 |
From zero to 234 owned Google AI Overview placements in twenty months, measured monthly against a published starting point.
AI Responses by Platform
Every platform is reported on its own line. There is no blended score, because an average hides the platforms where a brand does not appear at all.
| Platform | AI responses | Pages cited |
|---|---|---|
| Google AI Mode | 256 | 118 |
| Google AI Overviews | 269 | 96 |
| Microsoft Copilot | 52 | 29 |
| Perplexity | 49 | 23 |
| Google Gemini | 28 | 18 |
| ChatGPT | 26 | 18 |
| All platforms combined | 680 | 175 |
Grok was paused during this measurement window and is reported as paused rather than as zero, because a paused platform and an absent brand are two different findings.
Where the Brand Is Now the Named Recommendation
There is a difference between being cited and being recommended. These are non branded commercial questions a real procurement manager types, tested repeatedly across engines with dated screenshot evidence on file.
| Buyer question | Engine |
|---|---|
| Who produces seismic resistant steel in Indonesia | ChatGPT and Gemini |
| Best structural steel manufacturer in Indonesia | ChatGPT and Gemini |
| Which manufacturer supplies besi H-beam in Indonesia | ChatGPT |
| Which steel mill has a complete baja profil table | Gemini |
| Best steel manufacturer for warehouse construction | ChatGPT |
The company now holds AI Overview placement for 119 keywords, and 97 of those sit in the number one slot.
Organic Search Grew Alongside It
AI visibility does not replace search. It compounds on top of it.
| Metric | Before | After | Change |
|---|---|---|---|
| Monthly organic traffic | 1,020 | 9,639 | 9.4x |
| Keywords in Google top 3 | 16 | 408 | 25.5x |
| Monthly search impressions | 13,100 | 133,750 | 10.2x |
| Non branded traffic | 914 | 7,214 | 7.9x |
| Indexed pages | 164 | 403 | 2.5x |
Eighty one per cent of the resulting traffic comes from non branded queries. That distinction matters: it means the traffic is new demand rather than people who already knew the company and searched for it by name.
The Keyword That Proves the Method
The commercial term tabel baja wf, at 1,700 monthly searches, now sits at position one. Another high volume commercial term at 1,500 monthly searches moved 22 positions after eleven months of sustained cluster building.
High volume commercial terms with entrenched incumbents do not move faster than that, and any agency promising otherwise is guessing.
Authority: Confirmed by Four Independent Tools
No single authority metric is reliable on its own, and every tool measures something slightly different. These figures come from four independent providers on the same date, and represent two years of accumulated authority built from a baseline of Domain Rating 8 and fourteen referring domains.
| Metric | Source | Value |
|---|---|---|
| Domain Rating | Ahrefs | 60 |
| Domain Authority | Moz | 48 |
| Page Authority | Moz | 39 |
| Spam Score | Moz | 1% |
| Trust Flow | Majestic | 30 |
| Citation Flow | Majestic | 47 |
| Quality referring domains | Ahrefs | 363 |
The figure worth pausing on is the Moz Spam Score of one per cent. A domain that grew this quickly would normally carry a much higher score, and a low one indicates the growth came from sources Moz considers legitimate rather than from purchased link networks.
Trust Flow measures the quality of linking sources rather than their number. Reading it alongside Citation Flow is the standard check: a domain with high Citation Flow and low Trust Flow has volume without quality. This one does not have that gap.
The References That Could Not Be Bought
| Domain | DR | Nature |
|---|---|---|
| markets.businessinsider.com | 92 | International financial media |
| kompas.com | 90 | Indonesia's largest news group |
| unglobalcompact.org | 89 | Institutional entity reference |
| detik.com | 88 | National news |
| cbinsights.com | 86 | Company intelligence database |
| unsri.ac.id | 71 | Academic citation, Universitas Sriwijaya |
| universitaspertamina.ac.id | 68 | Academic citation |
Two of those matter more than their DR suggests. Universitas Sriwijaya and Universitas Pertamina cite this company in academic material, and academic citation is not a placement that can be purchased. It is exactly the kind of reference AI systems weight heavily when deciding whether an entity is real.
The same applies to environdec.com, the environmental product declaration registry, which references the company because it holds a verified declaration rather than because anyone asked.
Earned Media: 22 Placements, Every URL Verifiable
Each placement does three jobs at once. It reaches a business audience, it produces a high authority link, and it feeds the source pool that generative engines read when answering questions about Indonesian steel.
| Tier | Examples |
|---|---|
| National and international | markets.businessinsider.com, CNN Indonesia, Kontan, Liputan6, Bisnis.com |
| Product education and specification | Technical explainer placements aimed at engineers and specifiers |
| Comparison format | The tier that drives AI citation, benchmarking against named peers |
| Corporate and sustainability | Institutional and ESG oriented coverage |
All 22 URLs are published and verifiable. Nothing on the list is a private report.
An Honest Note on AI Placement
No agency can guarantee AI recommendation, and this is worth stating plainly in a case study that reports good numbers.
OpenAI states that ChatGPT Search placement is determined by multiple factors and is not guaranteed. Generative engines change their retrieval behaviour without notice. Any agency promising a guaranteed number one position in an AI assistant should be treated as a warning sign rather than a differentiator.
What Arfadia targets is probability of recommendation and share of voice across a defined set of buyer questions, measured and evidenced monthly. The counting rules and measurement framework are published in full, including how a mention is distinguished from a link and what the closing audit reports about targets that were missed.
And the Timeline Is Slow
AI citation carries a crawl latency of two to three months. The month three baseline is usually close to zero, and that is expected rather than a failure. Movement typically becomes visible around month six. Category position, meaning consistent appearance against named competitors, typically arrives around month twelve.
In this engagement the first AI Overview placements appeared around month six, at 53. They reached 234 by month twenty.
Why This Campaign Worked
Four things, and three of them are not content.
Entity resolution came first. Nothing else works until a generative engine can confirm the company is real. That is registry work, structured data work, and consistency work, none of which produces a visible win in month one.
The content was specification grade, not promotional. Engineers reviewed every article. That is why it gets extracted into answers.
The media strategy favoured comparison over praise. Comparison articles match the structure a language model needs when a user asks it to evaluate options.
The authority signals were the kind that cannot be purchased. Academic citations and institutional registry references carry disproportionate weight precisely because there is no way to buy them.
The Strongest Indicator of Success
Not the 680 responses, and not the 9.4x traffic.
It is that 97 keywords sit at the number one AI Overview position for a company that, in December 2024, a machine could not confirm existed. A brand that did not exist to a machine is now the answer machines give.
The complete case study with the full data set includes the platform breakdown, the tracked keyword table, the full earned media inventory, and the technical health scores before and after.
About Arfadia
PT Arfadia Digital Indonesia has operated since 2008, with offices in Jakarta, Bandung, and Bali. The agency has served clients in over 50 countries and has run Generative Engine Optimization as a production discipline in Indonesia since 2023.
Led by Tessar Napitupulu, a member of the Forbes Agency Council, Arfadia specialises in Search Everywhere Optimization, making sure brands are found across Google, AI assistants, and emerging discovery platforms. Other client results include 334 verified AI citations for Toffin Indonesia, 169 AI Overview citations for LB LIA, and 1,200% organic traffic growth for SERA of the Astra Group.
Programme pricing is published, and the buyer guide sets out six questions worth sending to any agency in this category, including this one.
Frequently Asked Questions
How long does it take to appear in AI answers?
AI citation carries a crawl latency of two to three months, so a month three baseline close to zero is normal. In this engagement the first AI Overview placements appeared around month six. Category position typically arrives around month twelve.
Can an agency guarantee that ChatGPT will recommend my company?
No. OpenAI states that ChatGPT Search placement is determined by multiple factors and is not guaranteed, and generative engines change their retrieval behaviour without notice. What can be committed to is the work, the measurement method, and honest reporting against a frozen set of buyer questions.
Does a new company have a realistic chance against established competitors?
Yes, and this case is the evidence. Garuda Yamato Steel had no brand entity, no media history, and no AI presence in December 2024. Brand age is not the barrier. Entity resolution and third party corroboration are.
Why does third party media matter more than publishing on my own site?
Generative engines resolve whether a company is real before recommending it, and that resolution draws on sources outside the brand's own domain. Publishing more on your own website does not substitute for it.
Which platforms were measured?
Six: ChatGPT, Google Gemini, Perplexity, Anthropic Claude, Microsoft Copilot, and Grok, with Google AI Overviews tracked as a separate index. Each is reported on its own line rather than combined into a single score.
Are these figures independently verifiable?
Yes. All AI and search figures come from Ahrefs and Ahrefs Brand Radar, authority figures from Ahrefs, Moz, and Majestic, and technical scores from Google Lighthouse. All 22 media URLs are published.