How to Choose an AI Visibility (GEO) Agency in 2026: Red Flags and Green Flags

AI visibility (also called GEO — Generative Engine Optimization, or AEO — Answer Engine Optimization) is the practice of raising the odds that AI answer engines cite your brand. The Princeton/Georgia Tech GEO study (Aggarwal et al., 2023) showed that adding citations, quotations, and statistics to content measurably improves its visibility — by up to 40% on their benchmark ^[4].

The important word is measurably. An agency that cannot say what it will measure, on which queries, and how often, is not doing GEO. It is doing SEO with a new label.

1. "We guarantee placement in ChatGPT answers." Nobody controls a model's output. Anyone guaranteeing citation placement is lying, or is selling you a definition of "placement" you would not recognize. The honest version: "we raise the probability, and we measure the result on a frozen query panel." Run.

2. No baseline, no panel. If the proposal jumps straight to activity (X articles, Y posts) without first measuring where you stand — citation share-of-voice, which queries matter, which competitors get cited — you are buying activity, not outcomes. A baseline is the only honest starting point.

3. Traffic metrics only. "We will grow your organic traffic" is an SEO answer to a GEO question. AI visibility is measured in citations, share-of-voice, sentiment, and assisted conversions. An agency that reports only sessions and rankings has not changed its model since 2019.

4. Proprietary "AI score" with no methodology. Several tools and agencies sell a black-box 0–100 score. If they cannot explain what the score counts, it exists to make the report look technical, not to make decisions better. Ask for the formula. Watch the pause.

5. Case studies without numbers. "We helped a leading SaaS brand improve its AI presence" tells you nothing. Real case studies show a baseline, a timeframe, and a delta — like "cited in 4 of 30 tracked queries in January, 11 of 30 in March." Vague adjectives are not evidence.

1. They show you a frozen query panel before they sell you anything. Good shops start by measuring: 20–50 queries your buyers actually ask, your citations, and your competitors'. This takes a day and costs nothing to show. If they will not do it, ask why.

2. They talk about content structure, not just content volume. Answer-first passages, entity consistency, original data, quotable facts. The GEO research is clear that how a passage is written changes whether engines cite it ^[4]. If the plan is "more blog posts," you are looking at an SEO agency wearing a GEO hat.

3. They are honest about what does not work. Ask "what have you tried that failed?" An agency with real experiments has a list. An agency selling a template has a pause.

4. Sentiment and accuracy are in scope. Getting cited is only half the problem — what the engine says about you matters. Good agencies track whether citations are accurate and favorable, and have a plan for when they are not.

5. They tell you when you do not need them. For some companies — purely local services, ultra-niche industrial B2B — AI-visibility spend is premature. An agency that says "your money is better spent elsewhere first" is the one to hire later.

Ask these verbatim. A measurement-first agency answers all five without reaching for a slide deck. A template agency redirects to "our proven process."

1. "Show me a before/after measurement from a real client — with the query panel."
2. "What exactly do you count as a citation, and how do you dedupe across engines?"
3. "What is your plan for the first 30 days — and what will the day-30 report show?"
4. "What did you try in the last year that did not work?"
5. "How do you handle engines citing a competitor for queries we should own?"

The third question is the sharpest. An agency that has never re-measured a frozen panel cannot describe what a day-30 report looks like, because they have never produced one.

Do not sign a six-month retainer as a first step. A sane pilot runs 4–6 weeks: baseline measurement on a fixed query set, 3–5 high-leverage content changes, a re-measure, and a report on what moved and why.

You should pay for the pilot — free pilots produce theater, not work. But it should be priced like an experiment, not a retainer. If the pilot doubles your citation share-of-voice, you have evidence to scale. If it does not, you bought a cheap, honest answer. That still beats six months of activity-based invoicing.

Here is what most agencies will not tell you. The content is not the hard part. The hard part is discipline: define the panel, measure the baseline, ship structured work, re-measure weekly, and tell the client the truth when nothing moved.

AI answer engines reward a specific set of signals — citations to credible sources, direct quotations, statistical grounding, and answer-first structure ^[4]. Those signals are repeatable. But they only compound if you re-measure on the same panel. Without a frozen query set, a "before/after" is just two unrelated screenshots. With one, a 0.9%→21.5% swing is a claim you can verify.

That is why we publish our own numbers below. Not because we are unusually good — because this is what evidence looks like.

"If an agency can't show you the query panel and the baseline before they invoice, you're buying confidence, not measurement." — Gregory Shevchenko, founder, Humanswith.AI

The difference between an agency that measures and one that performs is visible in their case studies. Below are our own published cases — every one has a baseline, a locked query panel, a timeframe, and a delta. We publish them because this is the standard we hold ourselves to, and it is the standard you should hold any agency to.

Full methodology and all eight cases: AI Visibility Case Hub.

[1] Gartner — "Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents" (press release, February 19, 2024) — https://www.gartner.com/en/newsroom/press-releases/2024-02-19-gartner-predicts-search-engine-volume-will-drop-25-percent-by-2026-due-to-ai-chatbots-and-other-virtual-agents

[2] Fractl — "Search Isn't Dying, It's Redistributing" (April 2026) — across 1,010,848 high-volume keywords, 29% of search volume is in measurable decline, surpassing Gartner's predicted shift — https://www.frac.tl/ai-search-study/

[3] Rand Fishkin / SparkToro — "In 2026, Less than One Third of Google Searches Still Send a Click" (2026 zero-click click-stream study) — https://sparktoro.com/blog/zero-click-search-study-2024/

[4] Aggarwal et al. — "GEO: Generative Engine Optimization" (Princeton / Georgia Tech / IIT Delhi, 2023, arXiv:2311.09735) — citations, quotations, and statistics improve generative-engine visibility by up to 40% — https://arxiv.org/abs/2311.09735