12 min read
Adapt by changing only what measured evidence shows is changing, and treat everything else as a small, time-boxed experiment.
In practice, that comes down to four moves. Keep your SEO and editorial foundations, because those are the page qualities most associated with AI citation. Invest more in earned, third-party coverage, because AI answer engines lean heavily on it. Measure beyond clicks. And give any new AI tactic a cheap test with an exit criterion before it earns a place in your strategy. That is enough to keep up with how buyers now discover brands without rebuilding your plan every time a new trick circulates.
Three shifts are real enough to act on
Most AI search advice mixes durable changes with speculation. The evidence clearly supports three shifts.
Answered queries send fewer clicks
When Google shows an AI summary, people click out less. In Pew Research Center field data on real searches, users clicked a link on 8% of searches that showed an AI summary, compared with 15% without one. Only about 1% of clicks happened inside the AI box itself, and roughly 26% of searches with a summary ended the session with no click at all (Pew Research Center data, as cited in Chen et al.).
Two qualifications keep this from becoming panic.
- Summaries appeared on only about 18% of observed queries. The click drop applies to those searches. It is not a site-wide traffic forecast.
- The size of the effect depends on context. A small controlled lab study of Google AI Overviews found no significant difference in engagement or website traffic, although people looked at the AI summary more (Rolon-merette, Guo et al.). Field data is stronger evidence of real behavior, but the lab result is a reminder to check your own query mix instead of assuming a fixed percentage loss.
This is also not a new problem. A causal analysis of 477,485 Google results pages found that direct answers and featured snippets were already lowering click-through rates before AI Overviews existed (Gleason et al., 2023). AI summaries extend a trend your team has handled before. When the answer sits on the results page, fewer people need to visit.
AI engines favor what others say about you
AI search engines show a strong bias toward earned media, meaning third-party, authoritative sources, over brand-owned and social content. Google's mix is more balanced (Chen et al.).
This matters for budgeting. Work on reviews, directories, industry publications, community discussions, and comparison articles now does double duty: it builds trust with human buyers and supplies the kind of sources AI engines tend to draw on. SEO, PR, and community work end up serving the same goal.
To decide where that effort goes, start by checking which domains the engines already cite when they answer your category's prompts. Prioritize getting included in those specific sources over broad outreach to every publication in your space.
Visibility is fragmented across engines
AI engines differ from each other in domain diversity, freshness, stability across languages, and sensitivity to how a question is phrased (Chen et al.).
An audit of 2,729 businesses across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews shows how uneven this gets. It found that 69.5% of businesses weren't mentioned by any of the five systems, and only 1.5% were mentioned by all five. Off-site signals such as Reddit and Quora mentions, Wikipedia, LinkedIn, and review sites had small positive associations with mentions in Claude, ChatGPT, and Perplexity, but not in Gemini or Google AI Overviews (MentionLayer, 2026). The audit is correlational, comes from an AI-visibility vendor, and hasn't been peer reviewed, so treat it as directional.
The practical takeaway is to cover several engines. You don't need a separate playbook for each one. You do need to stop judging your position from a single engine or a single answer.
Google isn't going away, so don't abandon search
AI chat is used at large scale. About 34% of U.S. adults report having used ChatGPT, and Perplexity disclosed 780 million monthly queries in May 2025. Google still holds roughly 90% of global traditional search. The same authors describe the change as "a steady reallocation of query resolution," not a replacement, and they acknowledge that their picture of the landscape is partly speculative (Chen et al.).
For planning, that means you should add AI channels to the plan rather than swap search out.
The foundations are what AI engines appear to reward
The page qualities most associated with AI citation look like ordinary technical SEO and editorial discipline. A study of B2B SaaS prompts across Brave, Google AI Overviews, and Perplexity found that metadata and freshness, semantic HTML, and structured data were the qualities most strongly linked to citation. Overall page quality was a strong predictor, with an odds ratio of 4.2 (Wrodium Research). The sample was small (70 prompts), the authors have a vendor affiliation, and the results are correlational. The direction is still useful: clear structure, accurate and current metadata, valid markup, answer-first writing, and sourced claims.
Foundations are necessary, but they aren't enough on their own. Overlap between AI results and Google results can be modest (Chen et al.), so ranking well doesn't automatically make you visible in AI answers. That gap is why the earned-media and multi-engine work above matters. For a closer look at how much of this is genuinely new, see Is AEO Really a New Discipline?
Why most AI tactics have a short shelf life
The tactics that spread fastest usually exploit how current models read text, which makes them fragile.
The research is real. The foundational GEO paper showed that content changes could raise visibility in generative engine answers by up to 40% on a research benchmark. It also found that results varied by domain, and it described these engines as black-box and fast-moving, with creators having "little to no control" over how their content is displayed (Aggarwal et al., 2024). "Up to 40%" is a best case on a benchmark, not a typical result on live engines.
Other studies show how manipulable model rankings can be. In controlled settings, inserting an optimized text sequence into a product page pushed that product to the top of a model's recommendations. Simple relevance tweaks, such as starting a passage with the query, beat credibility features like citations (Kumar & Lakkaraju, 2024; Wan et al., 2024, as cited in Chen et al.).
A tactic that works because of a model quirk works only until the model changes. Nobody has documented that engines will patch these specific exploits. But you would be building on a system you can't see, which changes without notice, and where the same trick may not carry across domains or engines.
A five-question filter for any new AI tactic
This filter is our own editorial framework, built from the evidence above. Before a tactic gets real budget, ask:
- Does it respond to a measured shift? Examples are fewer clicks on answered queries, the earned-media bias, and fragmentation across engines. A screenshot of one AI answer doesn't count.
- Would it improve the page for a human reader? Clearer answers, better sourcing, and fresher facts pass. Hidden text written for models fails.
- Does it plausibly work across several engines? Engines source differently, so a single-engine win is fragile.
- Does it work without relying on a model quirk? If the mechanism is "models over-weight X," expect it to decay.
- Can you test it cheaply and walk away? If you can't define what failure looks like, don't start.
A yes is the good answer to every question. Apply them in this order:
- Fails question 5: don't start.
- Fails question 2 or 4: a small, time-boxed experiment at most.
- Passes questions 1–4 and has a cheap test: run the test, and adopt the tactic only if it meets your exit criterion.
llms.txt is a useful example. It's cheap to add, and no evidence we reviewed confirms or rules out that AI providers use it. It fits in the "cheap test, low expectations" bucket. It isn't a strategy.
How to run an AI tactic test that won't hijack your roadmap
Write the test down before you start. A workable template:
- Hypothesis: "Adding a direct-answer summary and updated comparison data to our top five pricing pages will increase how often we're mentioned for pricing prompts."
- Engines: at least three, for example ChatGPT, Perplexity, and Google AI Overviews.
- Prompt sample: a fixed set of realistic buyer prompts, including varied phrasings, since engines are sensitive to wording. Run each more than once.
- Time window: long enough to see a pattern across repeated runs rather than a one-day swing.
- Exit criterion: "If mention rate doesn't move across at least two engines by the end of the window, we stop and revert the effort."
Pay attention to whether any change shows up across engines and phrasings, or only in one place. Consistent, cross-engine movement is a signal. A single good answer is noise.
Measure the chain, not just the clicks
If clicks are falling on answered queries, organic clicks alone will understate your reach. The fix is to track a ladder of signals and be clear about what each one proves:
Scroll horizontally to see all columns →
| Signal | What it tells you | What it doesn't prove |
|---|---|---|
| AI crawler visit | An AI system fetched your page | Indexing, citation, or a conversion |
| Mention or citation in an answer | You appeared for a prompt | A click, visit, or sale |
| AI referral visit | A person arrived from an AI source | That every AI-originated visit was captured |
| Conversion by landing page | Business impact | Which single answer caused it |
These distinctions come from how AI bot detection and referral tracking work. A bot request doesn't establish indexing or citation (Tideflow AI). Referrals depend on identifiable source information, so some AI-originated visits can't be recognized (Tideflow AI). And a mention count on its own isn't evidence of a click or a sale (Tideflow AI).
Judge visibility by patterns across many prompts and engines. Given that only 1.5% of audited businesses appeared in all five systems, a single snapshot tells you very little about what your buyers see.
Where a connected workflow helps
The hard part of this approach isn't knowing the principles. It's running the loop consistently: monitoring many prompts across several engines, spotting where competitors or third-party sources are cited instead of you, deciding whether to fix a page, write something new, or pursue inclusion in a source AI already cites, and then checking whether any of it produced crawler visits, referrals, and conversions.
Tideflow AI is built for that loop. It is currently in early access and working with design partners. It monitors prompts across ChatGPT, Gemini, Perplexity, Google AI Mode, and Google AI Overview. It turns gaps into specific plans and drafts content for human review. It publishes through MCP into your existing stack, logs AI crawler visits, and attributes AI referrals and conversions to landing pages. It is also explicit that neither monitoring nor publishing guarantees a citation, recommendation, or ranking (Tideflow AI). Its role is to publish the content your brand is actually missing, not to publish more.
Frequently Asked Questions
Does Domain Authority predict AI visibility?
Not reliably. In the audit of 2,729 businesses, Domain Authority wasn't a significant predictor of AI mentions once other factors were controlled for (MentionLayer, 2026). Treat it as an SEO heuristic, not an AI visibility metric.
Should we pursue third-party listicles and "best of" articles?
They can be worth pursuing, because earned media is where AI engines lean. Keep expectations measured, though. In the 2,729-business audit, off-site signals were linked to mentions in ChatGPT, Claude, and Perplexity but not in Gemini or Google AI Overviews, and the effects were small and correlational (MentionLayer, 2026). Prioritize relevant sources that engines already cite over volume. Are Third-Party Listicles Good GEO or Just AI-Era Link Building? examines whether they are real GEO or AI-era link building.
Does structured data help with AI answers?
Structured data was among the page qualities most associated with AI citation in a B2B SaaS study, alongside metadata, freshness, and semantic HTML (Wrodium Research). The finding is correlational, but valid structured data is cheap, helps traditional search too, and passes every question in the filter above.
If you want to put these principles into a concrete sequence, the next step is How to Get Mentioned by ChatGPT: A Tested, Step-by-Step Playbook.
Sources
- Chen et al., Generative Engine Optimization: How to Dominate AI Search
- Rolon-merette, Guo et al., Generative AI Summaries Do Not Increase User Engagement (2024)
- Gleason, Hu, Robertson & Wilson, Google the Gatekeeper: How Search Components Affect Clicks and Attention (June 2023)
- MentionLayer, Off-Page Signals Have Model-Specific Effects on Generative AI Search Visibility: Evidence from a Cross-Platform Audit of 2,729 Businesses Across Five Generative AI Systems (May 2026)
- Wrodium Research, AI Answer Engine Citation Behavior: Bringing the GEO-16 Framework in B2B SaaS
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, GEO: Generative Engine Optimization (August 2024)
- Tideflow AI, Detect AI Bots; Add analytics and track views and conversions; Tideflow AI vs Peec AI: Which AI Visibility Platform Should You Choose?; Tideflow AI vs Profound: Which AI Visibility Platform Should You Choose?