Instagram Beats OTAs in AI Travel Citations

By Tessar Napitupulu, Founder & CEO, PT Arfadia Digital Indonesia, GEO pioneer in Indonesia since 2023. More from Tessar.

In a July 2026 study of 105 Bahasa Indonesia travel queries run through Google AI Overviews, Instagram was cited 104 times. Traveloka was cited 72 times. Tiket.com was cited 33 times. A single social platform, made up almost entirely of individual operator accounts posting captions, out-cited Indonesia's two largest online travel agencies, individually and combined. Nobody built a strategy to make that happen. It happened because information-dense Instagram captions already look, to an AI system assembling an answer, exactly like a structured web page.

That single finding is the most consequential piece of evidence available today on how generative engine optimization actually works for Indonesian travel, and most tour operators have not heard of it yet.

What Exactly Did the July 2026 Study Find?

Search.Agency ran 105 Bahasa Indonesia travel queries through Google AI Overviews on 3 and 4 July 2026. Ninety-four of the 105 queries produced an AI-generated answer, a 90% trigger rate that confirms AI search is not experimental for Indonesian travel, it is operationally live right now. Those 94 answers generated 722 total citations spread across 298 unique domains, a spread wide enough that domain authority alone clearly does not decide who gets cited.

Several specific findings inside that dataset matter more than the topline numbers:

• Tour package queries triggered an AI answer 17 of 18 times, a 94% trigger rate, one of the highest of any query type tested. Package content is squarely inside the zone AI systems answer directly, not a marginal case.
• 69 of the 94 answered queries cited at least one social or user-generated-content source. Google's AI is treating social content as a first-class citation surface, not a supplementary one.
• Trust queries, specifically ones containing words like "amanah" or "terpercaya," produced the most consequential answers in the entire dataset: the AI named specific agencies, quoted price bands, and invoked official licensing as the evaluative criterion, while sourcing that verdict from the agencies' own marketing pages and Instagram reels rather than from the government body that actually issues the license.

The Itinerary Example That Explains the Mechanism

One example from the study makes the underlying mechanism concrete rather than abstract. A query asking for a Bali four-day itinerary returned an hour-by-hour plan with named locations, stated entrance fees, and itemized budget lines, "arrival 13.00, Pantai Melasti 14.30, Rp 5,000 entrance." The primary source the AI drew that answer from was Tiket.com's content-marketing itinerary article, not its booking page. The booking page was not cited for the planning query at all. The blog-style article, structured as a labeled day plan with real numbers attached to every element, was.

The study's own summary of the pattern across all 94 answered queries is worth quoting almost verbatim, because it describes exactly what to build: a short intro paragraph that commits to a direct answer, then a labeled list where each item carries its own citation, closing with two or three follow-up questions about budget, dates, or party size. AI systems are not extracting essays. They are harvesting labeled parts. A long, well-written page that is not shaped like the answer format the AI assembles can lose the citation slot to a thinner page that happens to be structured correctly.

Why Would an AI System Prefer a Caption Over a Booking Platform?

Because the caption, done correctly, contains exactly the kind of self-contained factual statement a retrieval system is designed to extract, and the booking platform mostly does not. The captions that earned citations in the study shared a consistent structure:

• A named package with an explicit price: "Paket Flores 5D4N mulai Rp 3.800.000/pax"
• Named included activities: "termasuk: open trip Kelimutu, Bajawa, Riung 17 Islands"
• Explicit departure dates: "Keberangkatan: 14 Juli, 28 Juli, 11 Agustus"
• An availability signal: "Sisa 3 seat"

None of that is complicated to produce. It is a discipline, not a technical build. A booking platform's real-time inventory page, by contrast, is optimized for a human filtering live search results, not for a retrieval system trying to lift a single quotable fact. The mismatch is structural, and it explains why a caption written by an individual operator can out-cite a platform with orders of magnitude more domain authority.

How Does This Fit Alongside Traditional SEO?

It doesn't replace it, it sits next to it, and the two have a fairly clean division of labor by query type.

Query type	Who currently wins the citation	Primary channel to build for
Bookable inventory ("hotel bali murah")	OTAs, structurally	Classic SEO, Google Hotels/Flights modules
Itinerary planning ("itinerary bali 4 hari")	Content-marketing articles, Instagram	Answer-shaped GEO content
Trust ("agen umroh terpercaya")	Operators' own pages, not government sources	Licensing and credential content
Destination inspiration ("tempat wisata terbaik")	High-authority editorial and government sites	Long-form destination guides, MaiA alignment

Mentions and Citations Are Not the Same Thing, and the Difference Has a Dollar Value

BrightEdge draws a distinction worth holding onto precisely: a mention names a brand inside the generated text, a citation attaches a clickable source link to that mention. Google AI Overviews cite roughly 3.2 sources per answer on average, and only a fraction of those citations overlap with the same domain's top-10 ranking in classic organic search. A brand can rank well and still be functionally invisible in the AI answer that increasingly sits above it, and the reverse is also true: a page with no ranking history can earn a citation purely on structural merit. Tracking only rankings, or only mentions without the linked citation, misses the metric that actually drives traffic.

How Contested Is This Ground Right Now, Realistically?

Less contested than most operators assume, which is exactly why the window is worth acting on. A SearchIntel audit of roughly 2,800 travel-brand queries across ChatGPT, Claude, and Gemini between March and May 2026 found an average AI-visibility score of 37 out of 100 across the category, with 90% of audited brands receiving zero citations at all on their core terms. A separate audit of Hawaiian excursion providers found an 82% correlation between the presence of structured schema markup and AI citation success, one of the more direct pieces of evidence tying technical implementation to citation outcome, though it comes from a single-vendor audit in a different market and should be read as directionally supportive rather than a universal formula.

Read together, those two numbers describe a field that is genuinely open and genuinely undefended. Ninety percent zero-citation is not a sign that citation is unattainable, it is a sign that almost nobody has tried the specific, low-cost changes the July 2026 Indonesian study shows actually work.

How Does MaiA Change the Picture for Indonesian Operators?

The Indonesian Ministry of Tourism launched MaiA, its own AI trip-planning assistant, on 28 November 2025, at the Sapta Pesona Building in Jakarta, under Tourism Minister Widiyanti Putri Wardhana, as the centerpiece of the government's "Tourism 5.0" program. MaiA is embedded directly in the Wonderful Indonesia website and generates personalized, real-time itineraries, dining suggestions, and destination recommendations, with a specific mandate to promote destinations beyond Bali, including Lake Toba, Borobudur, Mandalika, Labuan Bajo, and Likupang.

For an operator, MaiA is simultaneously a competitor and an opportunity. It is a government-backed AI itinerary builder that did not exist a year ago, and it is also a new, high-authority citation surface alongside ChatGPT, Perplexity, Gemini, and Google AI Overviews. Content that aligns with MaiA's own destination-diversification priorities, particularly well-structured content for the non-Bali destinations the government is actively promoting, stands to gain citation weight from a national tourism-authority source that most content on the web is not yet built to serve.

The Beyond-Bali Gap Is Real, and It Is Still Open

AI systems generating Indonesian travel recommendations default heavily toward Bali. This is a well-supported strategic inference rather than a directly measured statistic for Indonesia specifically, built from three separate strands of evidence: peer-reviewed research on AI destination concentration globally, documented anecdotal cases of AI itinerary generators defaulting repeatedly to Bali even when asked for alternatives, and the Indonesian government's own explicit framing of MaiA as an overtourism-mitigation tool. No study has yet directly measured AI mention rates for Flores, Raja Ampat, Toraja, or Lombok against Bali specifically. That gap in the evidence is itself the opportunity: first-mover, well-structured itinerary content for these destinations faces materially thinner competition for citation slots than anything written about Bali.

What Should an Operator Actually Do With This Information?

Treat Instagram captions as citation infrastructure, not just a follower-facing feed. Every caption for a real package should carry a named package, an explicit price, included activities, and departure dates, kept in sync with whatever the website says about the same package. Inconsistency between the two, a different price on the website than in the caption, is exactly the kind of entity-consistency failure the July 2026 study flagged as a source of AI errors, since the study found the AI "quotes with confidence, including its mistakes," and those mistakes trace back to inconsistent source content.

Second, publish the trust content the government source is failing to optimize. A dedicated, linked page stating TDUP or PPIU license numbers explicitly, BNSP guide certifications, and, for religious travel specifically, a "5 Pasti Umroh" style compliance checklist, directly answers the query type that produced the most consequential citations in the entire study.

Third, build itinerary content shaped the way the winning example was shaped: a direct-answer opening sentence, a labeled day-by-day plan with named locations and real prices attached to every line, and a closing FAQ section that pre-empts the two or three follow-up questions an AI answer typically surfaces on its own.

Schema Still Matters, It Just Isn't Sufficient Alone

TouristTrip schema with a fully populated itinerary property, structured as an ItemList of sequenced places, remains the highest-priority structured-data investment, followed by FAQPage on genuinely visible content, Organization or TravelAgency for entity clarity, and Certification schema for licenses specifically. None of this guarantees a citation on its own. What it does is make a page eligible for extraction at all, which a page without it structurally is not, regardless of how good the underlying writing is.

What Is the Realistic Measurement Approach Here?

Citation frequency in AI answers is genuinely harder to measure than a search ranking, because it drifts. Perplexity's citations for a given prompt drift roughly 40% month to month, and Google AI Overviews drift roughly 59%, which means a single test is a snapshot, not a stable baseline. The workable approach is a fixed panel of 15 to 25 representative prompts per destination and traveler profile, re-run weekly rather than monthly, tracking which competitor holds the cited slot when the operator's own content doesn't. That is reported through the RoGEO framework, which measures citation frequency, reference depth, and revenue attribution together, alongside the traditional SEO metrics that still govern bookable-inventory queries.

The strategic window here is not indefinite. The operators building this citation infrastructure now, before AI citation positions calcify around whichever early movers get there first, are the ones who will hold the advice-layer positions that compound going forward. Based on how quickly AI adoption in travel has moved over the past eighteen months, that window is realistically twelve to eighteen months wide, not permanent.

Does This Matter If Travelers Still Don't Trust AI to Book Anything?

It matters more, not less. A Skift survey published in March 2026 found that only 2% of leisure travelers are currently willing to let an AI system book a trip without a human reviewing it first. Read on its own, that number could suggest GEO is premature. Read alongside the citation data above, it says something more precise: travelers are using AI heavily for the advice layer, planning, comparing, and evaluating trust, while keeping a human decision point at the actual transaction. That is exactly the two-layer split this article has described throughout. GEO's near-term value is not booking automation, it is winning the advice conversation that happens before a human makes the final call.

That said, the infrastructure for the other layer is being built quickly elsewhere in travel. Priceline's "Penny" and pilots combining Sabre, PayPal, and Mindtrip point toward agentic, permissioned booking becoming real within the same platforms already doing the citing. An operator with clean, consistent, schema-backed data today is the one positioned to be selected when an agent does eventually get permission to transact on a traveler's behalf. An operator with messy or absent structured data will not become suddenly bookable by an AI agent just because the technology matures; the data foundation has to exist first.

The SEO side of this same strategy, winning the long-tail organic searches OTAs cannot cover at the same depth, is covered in our piece on why OTAs own generic search but not the booking, and the full service build is described on our GEO for travel and tour operators page.

---

Frequently Asked Questions

Is a 105-query study a large enough sample to build a strategy around?

It is a single, well-documented audit, not a longitudinal tracker, and should be treated with that scope in mind. It is direct, specific evidence rather than inference, which is genuinely rare in this field, but it describes one snapshot in time, not a permanent guarantee.

Does this mean we should stop investing in our website and focus only on Instagram?

No. The website remains where TouristTrip, FAQPage, and trust schema live, and where a traveler who does click through needs to land. Instagram is an additional citation surface with a specific, low-cost discipline attached to it, not a replacement for the site.

How often should we re-test whether our content is being cited?

Weekly, not monthly, given documented monthly citation drift of 40 to 59% depending on the platform. A monthly check will regularly mistake normal drift for a real change in performance.

Should our Instagram captions be in Bahasa Indonesia, English, or both?

Bahasa Indonesia is the dominant citation surface for Google AI Overviews specifically in the domestic Indonesian market. English content is separately needed to reach ChatGPT and Perplexity, whose indexed ecosystems remain largely English. Producing both, rather than translating one into the other, tends to perform better.

What is the single highest-priority first step if we are doing none of this today?

Publish one dedicated trust and licensing page with your actual PPIU or TDUP number stated explicitly, and rewrite your next ten Instagram captions to include a named package, explicit price, and departure date. Both are close to zero-cost changes with a direct line to the citation patterns this study documented.

If only 2% of travelers trust AI to book directly, why invest in GEO now?

Because the 2% figure describes trust in autonomous booking, not trust in AI-assisted planning, which is already mainstream. GEO wins the planning and advice conversation happening well before any booking decision, and the operators building clean, structured data now will be the ones ready when agentic booking does mature.

---