Augmented Visibility, How AI Recommendations Can Lead to Bad Business

AI assistants now shape which agencies a business owner calls first, and the evidence behind those recommendations is often thinner than it looks. In a March 2026 survey of 1,076 B2B software buyers and decision-makers, G2 found that 51% now start software research in an AI chatbot more often than in Google. 69% of respondents chose a different vendor than they had planned because of chatbot guidance, and a third of those switchers bought from a vendor they weren't familiar with before.

Those numbers come from software buying, but the habit they describe carries over to any vendor search, including the search for an agency. Trouble starts when an AI answer treats a paid ranking, a self-published "top 10" list, or a thin directory entry as independent evidence. The company named in the answer gets more visibility than its track record supports, and the buyer has no easy way to tell. We call that augmented visibility, and in a young category like AI search optimization it can push a business toward the wrong partner before anyone has checked what that partner has delivered.

What augmented visibility means

Augmented visibility is the gap between how prominent a company looks inside AI answers and how much independently verifiable experience sits behind that prominence. Ordinary visibility means a company is easy to find. Augmented visibility begins when an AI system lifts a company into a shortlist on the strength of content the company paid for or wrote about itself, then repeats that content in a neutral, confident voice.

Three things tend to blur together when this happens. A page can be findable, which only means the AI can retrieve it. It can be citable, which means it contains a clean claim the AI can quote. A company can also be recommendable, which means the AI puts it on the shortlist. None of those proves the company has run the kind of campaign you're about to pay for.

This is a separate problem from hallucination. In the cases below, the page exists, the claim on it is real, and the citation link works. What breaks is the step where a published claim gets read as proof of capability. It also differs from the gap we described when Gemini knew our company details but left us off a Jakarta SEO shortlist. That was about a capable company being overlooked, and this is about a shortlist filled on the wrong evidence.

How a paid ranking turns into an AI recommendation

The path from a promotional page to a shortlist is short, and context drops out at every step. OpenAI's own help documentation for ChatGPT search explains that it rewrites a question into one or more targeted queries, pulls web results, and builds the answer from them. The same page warns that "search results and citations can be incomplete, outdated, or incorrect." A favorable claim usually travels through that process in five steps.

1. A company or its PR vendor publishes a favorable claim, such as a "top GEO agencies in Indonesia" list, on its own blog, in a directory, or in a paid slot on a news site.
2. The page borrows credibility from its surroundings, including a familiar masthead, clean formatting, and a headline that sounds editorial.
3. A buyer asks an AI assistant a question worded almost exactly like that headline, so the assistant retrieves the page as a close match.
4. The assistant extracts the claim and leaves behind the context around it, such as who paid for the page, who wrote it, and whether the ranking has any method.
5. The buyer sees a tidy shortlist with a citation attached and reasonably assumes someone checked it.

Ranking pages fit step three almost perfectly, because a question like "best GEO agency in Indonesia" and a page titled "The 10 Best GEO Agencies in Indonesia" are near-identical, which helps explain why self-ranking lists have spread so fast. When Scrunch analyzed around 10,000 AI-cited URLs, about 30% of the comparison pages it examined were listicles in which the publisher ranked itself first.

What Google AI Mode returned for a GEO agency search in Jakarta

On 28 September 2026 we ran two buyer-style queries through Google AI Mode with the location set to Jakarta. One was in English ("best GEO agency in Indonesia") and one was in Bahasa Indonesia ("rekomendasi agency AI search atau GEO terbaik di Jakarta untuk perusahaan enterprise"). Both answers cited the same source among their top references. It was a ranking of GEO agencies for the Indonesian market, carried on a press-release feed hosted on a subdomain of a US community newspaper, and the same article is syndicated to other local mastheads.

The page is bylined by a content distribution firm and states "This is a paid placement." The Bahasa Indonesia answer carried the ranking's claims straight into its recommendation, including a growth figure for AI citations taken from the ranking. We haven't assessed that figure, and our point is only that the answer never said where it came from. The paid-placement label appeared nowhere in the answer, so a buyer reading it had no way to know one of its main sources was advertising without opening the link.

We're not naming the agencies on that list, and nothing here is a judgment on their work, since some of them may do excellent work. Our concern is the sourcing. The AI treated a page someone paid to publish as neutral evidence, for the exact question a marketing director asks when building a shortlist.

Search Agency appeared in that same Bahasa Indonesia answer when we ran it, and the sources behind our entry were our own About page and our LinkedIn company profile. When the same query was repeated later that day, the order of sources shifted and our entry dropped out, which shows how much these answers move between runs. We wrote both of those pages about ourselves, so any claim we make deserves the same checks as anyone else's.

The paid page getting through is less surprising once you see how rarely this kind of content gets cited in general. A BuzzStream and Citation Labs analysis of 3,600 prompts across ChatGPT, Gemini, AI Mode, and AI Overviews found that press releases distributed through syndication made up only 0.04% of all citations. Syndicated content usually struggles to get cited. Google's explainer on AI Overviews notes that for uncommon searches there might not be "a lot of high quality information available on the web," and a niche query like "GEO agency in Indonesia" appears to fit that description.

Why a well-known publication does not vouch for every page it hosts

Indonesia already has a rule meant to keep readers from confusing advertising with journalism. The Dewan Pers Pedoman Pemberitaan Media Siber, ratified in February 2012, requires online media to separate news from advertising and to label any paid content with a term such as "advertorial," "iklan," "ads," or "sponsored."

The paid page in our test did carry its label, so the gap sits in how AI answers reuse content rather than in the publisher's disclosure. A label protects a reader who is looking at the page. It does much less once an AI assistant has pulled a sentence off the page and dropped it into an answer, because the disclosure usually sits in a byline, a footer, or a design element that never makes the trip. Labels are easy to miss even on the original page. In a University of Georgia experiment published in the Journal of Advertising, only 17 of 242 participants, about 7%, recognized native advertising as advertising.

A strong domain can host newsroom reporting, partner content, sponsored rankings, and press-release feeds side by side, each held to a different editorial standard. The masthead's reputation belongs to the newsroom, and it tells you nothing about who picked the agencies on a paid list or whether anyone checked their results. Google draws the same distinction. Its August 2026 update to the site reputation abuse policy says action is likely when third-party content lacks author or editor identification, omits commercial disclaimers, and sits apart from a site's editorial sections. Most of what AI cites about any brand lives on sites the brand doesn't control, so the question for a buyer is who wrote a given page and who paid for it.

Why citations make AI answers feel more trustworthy than they are

A citation link makes an answer feel checked, even when nobody checked it. Researchers Haiwen Li and Sinan Aral ran a preregistered experiment with 4,927 US adults and found that adding reference links raised both trust in AI search results and willingness to share them. The increase was the same whether the links were valid or invalid. People who trusted AI more also clicked more and spent less time evaluating what they read.

The citations themselves are often weaker than they appear. A 2023 Stanford audit of four generative search engines (Bing Chat, NeevaAI, Perplexity, and YouChat) found that only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to. The authors called this a "facade of trustworthiness." In 2025 the Tow Center for Digital Journalism tested eight AI search tools on 1,600 source-attribution queries, and collectively they answered more than 60% incorrectly.

Neither study looked at agency selection, but together they describe the conditions a buyer works in. The answer sounds certain, the link underneath looks like proof, and the more a reader trusts the tool, the less likely they are to open that link.

How easily one page can move an AI shortlist

Researchers have now measured how much a single seller-controlled page can change what an AI recommends. The SafeGEO study, published in June 2026 by researchers at the University of Toronto and collaborators, tested 22 manipulation techniques across 600 recommendation cases in six product categories. Each product's real quality stayed fixed while the researchers rewrote one source the seller controlled.

Those rewrites raised a flawed product's appearances in the top three recommendations by up to 83.2% relative to baseline, and increased top picks that broke a user's stated requirements by up to 59.3%. One of the techniques tested is what the authors call authority laundering, where content tied to the seller is presented as independent guidance. A paid "best agencies" ranking on a news domain fits that description closely.

A separate benchmark called FORGE tested 12 models on 225 products and found that one polluted web page fooled the recommender up to 27% of the time. Manipulation also reaches past the open web. In February 2026, Microsoft's security team reported 50 hidden prompts from 31 companies in more than 14 industries, planted in "Summarize with AI" buttons to tell assistants to remember a company as a trusted source or to recommend it first.

Visibility gained this way also moves around. When Ahrefs published 34 self-promotional pages as an experiment, the citations showed up on only about one in three eligible days, and 43% of the answers that cited its conference pages never mentioned the conference. Lily Ray found a similar pattern in Google AI Overviews, where brands' own "best of" lists were cited while the brand was left out of the recommendation 69% of the time. A shortlist built on this kind of content can look different next week.

What an AI shortlist cannot tell you about an agency

An AI answer can tell you that a publication called an agency one of the best. It can't tell you whether that agency has run an AI search campaign, measured the result fairly, or kept a client long enough to see the work pay off, and those are the things a hiring decision rests on.

For an AI search engagement, the useful questions are specific. Has the agency measured a fixed set of prompts before and after its work, across more than one AI platform and in both Bahasa Indonesia and English? Can it separate its own effect from seasonality, PR, paid media, and ordinary SEO? Will a client confirm the engagement and the result on a call? Can it show a month where the numbers went down and explain why?

The field is young, which makes these questions harder to answer. The term "generative engine optimization" only entered use with a research paper published in November 2023, so few agencies anywhere have a long record in it, and most of them, including us, came to it from SEO. That's normal for a new discipline, and it's the reason a shortlist needs checking.

Getting the choice wrong costs more than a retainer. Months spent with a partner who can't deliver are months a competitor spends earning the citations and brand mentions AI systems draw on. There's a risk to your own brand as well, because Google's spam policies now state that they apply to generative AI responses in Search. Any agency should be able to tell you which tactics it would use on your behalf and how they fit those policies. G2 found that 64% of B2B buyers already run into inaccurate AI recommendations "often or very often," so many buyers know the answers can be wrong. Fewer have a routine for checking them.

How to check an AI-recommended agency before you sign

Use an AI recommendation to build a longlist, then do the due diligence yourself. Most of it takes an afternoon and a few direct questions. The table compares evidence that holds up with evidence that tends to fill paid rankings and sales decks, and the checklist after it covers most of the remaining risk before a first pitch meeting.

1. Open every source the AI cited for the agency and look for "advertorial," "sponsored," "paid placement," or "iklan" labels, including in the byline and footer.
2. Check who published each ranking, and whether the agency in first place also owns the site.
3. Search the agency's name to see where its awards and rankings were first published.
4. Ask the AI assistant which of its sources were sponsored or written by the companies it recommended, then verify the reply yourself.
5. Ask the agency for one AI search case study with a named client, a baseline, a timeframe, and the prompt set it measured.
6. Ask for a client reference and make the call.

Our playbook for vetting Indonesian B2B SEO agencies covers the commercial side of that conversation. Once an engagement starts, evaluating SEO work without an in-house specialist explains how to read the results, and measuring AI share of voice shows what a credible AI visibility report contains.

Where honest AI search work draws the line

AI search optimization makes a company easier for AI systems to find, understand, and quote. Done properly, that means publishing verified case studies, keeping company information consistent across the web, making pages easy for crawlers to read, and earning coverage from people with no stake in the outcome. The line is crossed when the work stops making real evidence easier to find and starts manufacturing the look of agreement, through self-ranked lists, undisclosed paid rankings, padded claims, or hidden instructions aimed at the model.

We publish how we measure AI search work in our methodology. It explains that we build the prompt set with each client and then freeze it, because swapping prompts mid-engagement resets the trend line. It also commits us to telling a client when a category has no meaningful AI answer volume and the budget belongs somewhere else. If you're comparing agencies, us included, ask each one to walk you through its method at that level of detail, and to show you a result that didn't go the way it planned.