Search Visibility and SEO in the Age of Generative AI - An evidence informed review and practical adaptation framework

Quick Takeaways

AI is changing search, but making your website useful and easy to find still matters. People can now get answers without visiting a website. That makes it more important to understand what brings readers to your pages and what helps them once they arrive.

• Being mentioned is not the same as getting results. An AI tool can cite your page without sending visitors. It can also misunderstand what you wrote.

• Start with real questions. Find out what your audience needs to understand, compare, or decide. Build your content around those needs.

• Give clear, trustworthy answers. Explain the main point early. Support important claims with reliable sources, and keep limitations beside the claims they qualify.

• Check that search systems can read your pages. Useful information cannot help discovery if technical problems prevent access.

• Improve a few pages first. Record their current performance, make a focused change, and compare results over time.

• Measure what matters to your website. Track accurate citations and useful visits alongside actions such as subscriptions, purchases, or use of a helpful resource.

You do not need a secret formula for AI search. Begin with accessible pages, useful answers, and evidence you can check. Then test whether your changes help readers and justify the work.

Abstract

Generative AI is changing how people discover information and how websites receive value from search. Evidence supports lower click propensity for some searches displaying AI summaries, but does not establish the disappearance of search engine optimization or a universal replacement of traffic by citations. This narrative review integrates independent browsing research, original industry datasets, peer reviewed information retrieval research, recent 2026 studies, and current platform documentation. It distinguishes search visibility, source citation, accurate representation, referral visits, and downstream outcomes. These are related but noninterchangeable measures. Findings differ by sampling frame, query intent, device, platform, and observation period. Emerging optimization experiments show that content changes can affect visibility within particular systems, while newer research exposes tradeoffs between retrieval and answer inclusion. A practical framework, supported by an execution workflow and worked example, combines technical accessibility, useful and verifiable content, repeated platform specific measurement, and controlled evaluation of business outcomes. The central implication is that SEO needs a broader measurement and investment model, rather than abandonment or a guaranteed formula for AI citations.

Keywords: AI search optimization, SEO, generative engine optimization, GEO, AI Overviews, AI search visibility, AI citations, organic traffic, technical SEO, content optimization, referral traffic, search analytics

How to use this article

If you are responsible for a website, begin with Section 5 and use Section 6 to define success. Sections 3 and 4 explain the evidence behind those decisions. You do not need to rebuild the site or purchase a monitoring platform to start: select a small group of useful pages, record current performance, check access, and test a focused improvement.

1 The question has changed

The useful question is no longer simply whether SEO is dying. It is which kinds of discovery still produce visits, which generate exposure without a visit, and which justify the cost of creating and maintaining content. A publisher supported by advertising, a local service business, and a software company may face similar interface changes but very different economic consequences.

This article uses stratification to describe how search value may become more unevenly distributed across queries, publishers, and forms of visibility. Citation exposure does not necessarily compensate financially for lost traffic: an AI answer may cite a page, summarize it incorrectly, and send almost no readers to it.

SEO and generative engine optimization, or GEO, are best treated as overlapping practices. Their relationship cannot be reduced to traditional keyword matching on one side and semantic AI retrieval on the other. Contemporary search and generative answers can share discovery infrastructure and sources while presenting information differently. The strategic task is to identify the bottleneck that prevents useful content from producing a worthwhile outcome.

2 Scope and evidence standards

This is a targeted narrative review, not a systematic review, meta analysis, or original empirical study. Searches and source checks were conducted through October 22, 2026. Priority was given to original research reports with identifiable methods, scholarly publications and clearly labeled preprints, and official documentation for implementation requirements. Recent publication dates alone were not treated as evidence of quality. Unverified bibliographic entries and secondary reports that obscured the underlying dataset were excluded.

The sources answer different questions. Browsing panels describe observed user behavior. Search Console studies characterize sampled impressions and clicks. Laboratory benchmarks test interventions within specified pipelines. Platform documentation defines eligibility and controls, rather than proving traffic benefits. Company deployment reports illustrate specific implementations. None alone identifies the total causal effect of generative AI on the open web.

Comparisons below retain their original units and periods. A visit level click probability is not a first position CTR; a modeled relative loss is not a percentage point change; and growth from a small referral base does not reveal total market share. These distinctions are essential to interpreting apparently contradictory findings.

3 What the traffic evidence supports

3.1 Independent browsing evidence

Chapekis and Lieb (2025) analyzed browsing from 900 U.S. adults, covering 68,879 unique Google searches in March 2025. Traditional result links were clicked on 8% of visits with an AI summary, compared with 15% without one. Links within summaries were clicked on 1% of visits with a summary. These figures describe visits, not the CTR of the first organic result.

The study reconstructed search pages in April from queries observed in March. Results could have changed between the actual visit and reconstruction. Summary exposure was also not randomized: longer and question shaped queries were more likely to produce summaries. The study therefore supports an association with lower clicking, not an isolated causal estimate of the summary interface (Chapekis & Lieb, 2025).

3.2 Large keyword samples and modeled losses

Law (2026), reporting Ahrefs research, compared 150,000 keywords with AI Overviews and 150,000 informational keywords without them using aggregate desktop Search Console data for December 2023 and December 2025. The widely cited 58% estimate compares observed first position CTR with an estimated counterfactual adjusted for the decline in the comparison group. It is not a measured 58 percentage point loss or a universal reduction for every website.

The comparison improves on attributing all change to AI Overviews, but requires the comparison group to approximate the trajectory the exposed group would otherwise have followed. Differences in query composition, competing interface features, and ranking or audience changes remain potential explanations. Its desktop scope also limits direct application to mobile traffic (Law, 2026).

3.3 Newer evidence complicates a continuous decline narrative

Seer Interactive’s April 2026 update covers 53 brands, 5.47 million tracked queries, and 2.43 billion organic impressions from January 2025 through February 2026. Aggregate organic CTR for queries labeled as showing an AI Overview rose from 1.31% in December to 2.36% in February; the comparison without an Overview reached 3.82%. This sample does not support an uninterrupted downward trajectory (Seer Interactive, 2026).

Important limits remain: Seer retrospectively applied a recent, static AI Overview label across the observation period. Its expanded sample differs from earlier reports, making direct splicing inappropriate. These figures also measure organic CTR for query groups, not clicks specifically on citation links. Its finding that impressions can rise while CTR falls reinforces the need to inspect absolute clicks alongside ratios (Seer Interactive, 2026).

Table 1  Selected traffic findings and their proper interpretation

Taken together, these studies justify planning for click pressure without assuming that every segment is shrinking at the same rate. The table is an evidence map, not a pooled effect estimate. Zero click behavior should likewise not be treated as synonymous with AI substitution: it can reflect reformulation, navigation within a platform, abandonment, or satisfaction without an external visit.

3.4 Referral quality is promising but conditional

Adobe’s August 2026 report provides newer sector evidence. In July, AI referred U.S. retail visits grew 62% year over year and converted 60% above non AI visits. For travel, AI referrals grew 119%, while conversion remained approximately 1% below non AI traffic. The contrast matters: superior conversion is not a universal characteristic of AI referrals (Pandya, 2026b).

An earlier Adobe report placed the retail conversion advantage at 42% in March 2026 (Pandya, 2026a). These changing figures describe observed cohorts, not the causal effect of steering a visitor through AI. Referral users may already have different intentions. Growth and conversion ratios also do not reveal whether additional profit offsets traffic lost elsewhere. A business needs its own channel volumes, margins, acquisition costs, and retention data.

4 How generative visibility should be understood

4.1 A pipeline with several possible failures

Lewis et al. (2020) established retrieval augmented generation as an architecture combining model parameters with retrieved information. That foundational system does not document the internal operation of every current commercial product. Dense retrieval, lexical matching, reranking, query decomposition, and other techniques may be combined differently; a single cosine similarity diagram cannot establish how a platform selects sources.

A useful diagnostic model separates access and indexing, candidate retrieval, selection into model context, answer generation and attribution, user response, and eventual value. A page can succeed at one stage and fail at the next. Retrieval does not ensure citation, and a visible citation does not ensure faithful representation. Because source selection depends on systems obtaining usable content, technical audits begin with access and indexing before testing improvements to retrieval or citation.

The newer SAGEO Arena study explicitly evaluates retrieval and reranking as well as generation. Its August 2026 revision, listed as accepted at KDD 2026, reports that existing approaches often degrade earlier pipeline stages and that structural information can mitigate some limitations. This strengthens the case for testing the entire discovery path instead of evaluating only content already supplied to a model (Kim et al., 2026).

4.2 What foundational GEO experiments actually establish

Aggarwal et al. (2024) showed that modifications including quotations, statistics, and citations could increase visibility in their experimental generative engine. The often repeated improvement of up to approximately 40% concerns defined answer visibility metrics, including position adjusted word count, rather than website traffic or sales. Fluency and readability changes also helped in their experiments, contradicting the idea that prose quality had little effect.

These findings justify hypotheses for testing. They do not guarantee that adding statistics will improve organic discovery across current platforms. Effects depended on the metric, domain, and starting source position. The editorial lesson is to provide relevant, accurate evidence and clear explanations, rather than manufacture numerical density or add citations merely for appearance (Aggarwal et al., 2024).

4.3 Production evidence and its limits

Zhang et al. (2026) describe Pinterest’s production approach using vision language models, inferred search intent, collection pages, and internal distribution. Their preprint reports 20% organic traffic growth. It illustrates a substantial engineering intervention beyond paragraph rewriting. However, its outcome is not an isolated estimate of AI citation gains, and its scale and visual corpus limit transfer to a small publisher. Conventional search improvements and generative visibility cannot be cleanly separated from that headline result.

4.4 Citations are neither proof nor payment

Liu et al. (2023) audited four early generative search engines and found that only 51.5% of generated sentences were fully supported by citations, while 74.5% of citations supported the associated sentence. These historical figures should not be presented as 2026 error rates. They establish why citation presence and citation accuracy must be evaluated separately.

For a publisher, visibility can therefore be beneficial, neutral, or harmful. A prominent but inaccurate attribution can misstate a product, recommendation, or research result. Quality monitoring should ask whether the answer represents the source correctly, as well as whether the domain appears. Economic evaluation must then determine whether the exposure produces a useful action.

5 A practical adaptation framework

AI search optimization means improving the chance that useful content is discovered, represented accurately, and connected to an audience outcome. The framework below combines three kinds of guidance: documented platform requirements, editorial practices that improve usefulness, and experiments whose visibility effects must be measured. GEO remains the term used when discussing the research literature.

5.1 Diagnose the affected work before changing strategy

Start with a page inventory organized by audience purpose and business contribution. Separate branded navigation, local discovery, general explanation, comparison, troubleshooting, and transaction tasks. For each group, inspect impressions, clicks, ranking changes, conversions, and publishing costs. Keep device, geography, seasonality, and major site changes visible. A falling CTR with stable clicks calls for a different response from collapsing demand or lost indexation.

This segmentation prevents a common error: applying a headline loss from informational desktop queries to all content and all revenue. Preserve pages that continue to help users or support other pages even when their direct attribution is modest.

5.2 Fix access and eligibility with platform specific controls

For Google, indexed pages eligible for a search snippet can qualify for supporting links in AI Overviews and AI Mode; no additional technical requirement or special AI schema is specified. Google also describes query fan out, in which a system issues related searches across subtopics to develop an answer (Google Search Central, n.d.-a). To assess topic coverage, list the related questions a reader needs answered and check whether the site addresses them clearly. This tests content coverage; it does not reveal the platform’s internal queries.

Search access must be distinguished from training permission. OpenAI documents OAI-SearchBot for search and GPTBot for potential model training, with independent controls; ChatGPT-User performs user initiated actions. A site can allow search access while declining training access (OpenAI, n.d.). Google Extended does not determine inclusion or ranking in Google Search (Google, n.d.). A recommendation to unblock all AI bots would therefore be unnecessarily broad.

Begin the technical audit with representative pages from each important template, including an article, product or service page, and interactive resource where applicable. Record the URL, test date, finding, and proposed remedy for each of the following checks.

1. Check the initial HTML. In browser developer tools, open Network, reload the page, select the main document request, and search its Response for a distinctive sentence from the main answer. Compare it with the rendered page in the Elements panel. If the answer appears only after scripts execute, document the rendering dependency before deciding whether server rendering or static HTML is needed.

2. Check Google’s rendered view. In Search Console’s URL Inspection tool, run a live test and inspect the tested page’s HTML, screenshot, and resource errors. Confirm that the main answer and essential links are present. Google can render JavaScript, so absence from the initial HTML alone does not establish invisibility to Google; test other crawlers separately (Google Search Central, n.d.-c).

3. Check access and indexing controls. Inspect status codes, redirects, robots.txt rules, noindex directives, canonical URLs, internal links, and CDN challenges. A successful response containing a challenge screen is not usable article content. Record whether each issue prevents fetching, obscures the answer, or affects indexing eligibility.

4. Verify requests in server logs. Check the source IP, requested URL, timestamp, response status, and response size. Verify Googlebot against Google’s published IP ranges or its reverse and forward DNS procedure; do not trust the user agent alone. Apply each other provider’s documented verification method before changing access rules (Google Search Central, n.d.-d).

Resolve observed access defects first, then retest the affected URLs. Preserve access controls for sensitive, private, and paid material. A working page template usually warrants a targeted repair rather than a wholesale framework migration.

5.3 Improve usefulness and verifiability

The editorial objective is a page that answers a real question and lets a reader check consequential claims. State the answer clearly, then explain conditions, exceptions, and supporting evidence. Use descriptive headings and actual HTML tables for comparisons when appropriate. Choose answer length to fit the question and test reader usefulness rather than imposing a fixed word count.

Evaluate content usefulness by comparing the page with the sources readers already encounter. Identify its specific contribution: a documented test, maintained tool, local knowledge, transparent comparison, or synthesis that resolves conflicting evidence. This added value is the practical meaning of information gain here, not a claimed proprietary ranking score. Check that consequential numbers have a denominator, collection date, method, and source. Improve unclear or unsupported passages before adding more content; originality need not require an expensive proprietary dataset.

Google frames experience, expertise, authoritativeness, and trustworthiness as qualities its systems seek to identify, while explaining that E E A T itself is not a specific ranking factor. Apply those ideas through accurate authorship, relevant credentials, source transparency, correction practices, and content suited to the reader’s purpose. Do not invent measurable trust vectors or assume that AI assisted authorship alone establishes low quality (Google Search Central, n.d.-b).

For CupTalks or another mental health publisher, this means distinguishing research findings from practical suggestions, giving study populations and limitations where relevant, and explaining whether evidence concerns symptoms, diagnosis, or treatment outcomes. This is an editorial application of the framework, not a claim that a particular formatting choice guarantees AI visibility. Reader understanding remains the primary outcome.

5.4 Give visitors a worthwhile next step

If an answer interface can satisfy a simple factual request, a website visit needs a clear purpose. That purpose might be a comparison tool, a detailed method, a maintained resource, a booking, or an optional subscription. The next step should follow naturally from the page’s subject. Withholding the basic answer merely to force a click undermines the usefulness that makes the source worth discovering.

5.5 Run a focused AI search optimization pilot

Step 1. Choose the audience task and outcome. Select a question that matters to readers and the site’s purpose. Define the useful action before editing: for example, completing a comparison, downloading a relevant resource, booking an appointment, or returning to a maintained guide. Choose a small set of related pages with enough existing activity to evaluate; use 30 to 50 audience questions for the initial prompt panel described in Section 6.

Step 2. Establish what happens now. Export page level search clicks, impressions, and useful actions for a period long enough to capture the site’s usual variation. Record the date range and major campaigns or site changes. Run the selected questions on the platforms your audience uses, repeat them over time, and preserve the answers, cited URLs, and any incorrect descriptions. Label an unobserved retrieval step as unknown rather than inferring it from a citation.

Step 3. Identify a specific gap. Compare cited pages and your own page against the reader’s task. Does the page answer the question directly? Does it explain the relevant conditions? Does it offer a source supported comparison or example? Competitor citations are leads for investigation, not proof that copying their format will work. Prioritize gaps that would matter even if no AI system cited the page.

Step 4. Make one coherent improvement. Use the technical checks in Section 5.2 for access problems. For a content gap, rewrite the opening answer, add the missing comparison or evidence, and keep qualifications beside the claim they limit. Add relevant internal links so users can find the page. Record exactly what changed; avoid simultaneously changing the URL, page design, title, and substantive content when the aim is to learn which intervention helped.

Step 5. Verify the published result. Confirm that the live page shows the intended answer and that the relevant crawler can obtain usable text. Check dates, authorship, citations, and consistency between visible content and any structured data. Fix inaccurate metadata; adding markup that does not match the page is not a substitute for better content. Retest a sample of the prompt panel after the page has had time to be discovered.

Step 6. Decide whether to expand. Compare the changed pages with appropriate unchanged pages using the approach in Section 7. Review accurate citations, search clicks, qualified referrals, useful actions, and maintenance cost together. Expand an improvement when the evidence supports the chosen outcome; revise it when it helps visibility but creates misunderstanding or loses reader value.

Use one audit record per page: URL; audience question; intended action; baseline dates and counts; access finding; content gap; planned change; owner; publication date; comparison pages; review date; and result. A spreadsheet is sufficient. Preserve the original page and change log so a disappointing change can be investigated or reversed.

5.6 Worked example of an evidence based answer

Consider a publisher improving a page for the question “Does AI search traffic convert better?” This example illustrates an editorial intervention, not a demonstrated citation gain. The intended audience is a website manager deciding whether to invest in the channel; the useful next action is to calculate the site’s own referral value.

A weak opening would be: “AI search is transforming digital marketing. Brands must embrace the future and optimize for AI to unlock high quality traffic.” It gives the reader no usable comparison, evidence, or decision rule.

A stronger opening would be: “AI referred visitors can convert better, but results depend on the sector and period. Adobe reported that U.S. retail AI referrals converted 60% above non AI traffic in July 2026, while travel AI referrals converted approximately 1% below it. These are observed differences, not proof that routing visitors through AI causes more purchases. Evaluate your own visits, conversion rate, and value per conversion before increasing spending” (Pandya, 2026b).

Follow that answer with a compact comparison showing sector, observation period, relative conversion difference, and source. Explain the denominator: “60% higher” is a relative difference, not a 60 percentage point increase. Link the underlying report, identify the update date, and explain what the data cannot establish. Keep this context near the numbers so an excerpt remains interpretable.

Then give readers a task they can complete: identify AI referrals in their analytics, count conversions under a stated definition, calculate conversion rate as conversions divided by eligible sessions, and compare the same period with relevant channels. Use contribution margin rather than revenue when assessing profitability, account for costs consistently, and inspect small counts before interpreting large percentage differences. Apply the worked deficit calculation in Section 6 to test whether higher conversion could plausibly offset lost visits.

The substantive improvement is a clearer, checkable answer tied to a decision. Whether it also increases citations is a separate outcome to test. A medical publisher can use the same editorial structure, while selecting appropriate clinical sources and obtaining relevant expert review for consequential health claims.

5.7 Match the remedy to the observed problem

Table 2  Practical diagnosis and first action

Prioritize a verified blocker affecting important pages before experimenting with formatting. Next address inaccurate or outdated consequential claims, then missing answers and comparisons. Use expensive monitoring or sitewide redesign only when the expected benefit justifies the cost. A failure to appear in a small prompt panel is not enough to diagnose an indexing problem.

5.8 Separate requirements from experiments

Documented requirements concern eligibility and access, such as Google’s indexing and snippet conditions and the independent search and training crawler controls described above. Clear explanations, traceable evidence, and useful comparisons are editorial practices. Particular paragraph lengths, question based headings, and statistics additions intended to improve citation rates are optimization hypotheses to test. Google specifies no special AI schema or AI text file requirement for its AI search features (Google Search Central, n.d.-a).

Do not manufacture quotations, statistics, credentials, independent reviews, or endorsements to look authoritative. Avoid hidden instructions aimed at manipulating AI answers. Audit whether the evidence genuinely supports the claim, and update content when the underlying facts change. The practical objective is to become a useful source that can be checked, with measured benefits that justify ongoing work.

6 Measure outcomes without confusing them

Maintain separate measures for being found, being represented, receiving a visit, and delivering value. A domain mention without a link is a mention, not a citation. A citation is an attribution, not proof that the source materially shaped the answer. Referral sessions are observed visits, not a complete estimate of all influence. These distinctions should appear in reporting definitions.

Table 3  A minimum measurement specification

The denominators in Table 3 are proposed reporting conventions. Report counts alongside percentages. Keep engines, languages, locales, devices, model versions when visible, and collection dates separate. A coverage estimate from a hand selected prompt panel should not be labeled market share or weighted by invented search volumes.

Begin with 30 to 50 questions drawn from real customer or reader needs, repeat them over several weeks, and expand coverage as resources permit. Include paraphrases, record whether search occurred, preserve outputs and cited URLs, and manually audit a subset for factual fidelity. Repeated outputs for the same prompt are correlated; thousands of runs do not automatically provide thousands of independent observations.

Use Search Console alongside analytics and server logs. Google’s documentation describes AI feature traffic within overall Web performance reporting, so the aggregate alone should not be interpreted as a clean citation traffic measure (Google Search Central, n.d.-a). Track identifiable external AI referrals separately, but retain established search and conversion reporting. An increase in direct traffic or branded queries is corroborating evidence of possible influence, not proof of AI attribution.

An illustrative calculation shows why conversion premiums need context. Losing 1,000 visits that previously converted at 2% loses 20 conversions. Gaining 100 AI referrals converting at 4% restores four. Even a doubled conversion rate leaves a deficit of 16, before differences in margin or repeat use. These are hypothetical numbers, not empirical estimates.

7 Test changes before scaling investment

A practical evaluation begins with one intervention and one primary outcome. For example, add a documented comparison and source supported explanation to a group of pages, retain comparable pages unchanged, and specify the observation window before checking results. Track both retrieval or search performance and answer visibility so that a gain at one stage does not conceal a loss at another.

Where practical, randomly assign eligible pages or content clusters. Otherwise match comparison pages on prior trends, intent, and traffic, and acknowledge that the estimated effect depends on those assumptions. Log concurrent site edits and platform changes. Pages on the same site can influence each other through links and authority, so controls are imperfect. Evaluate durability after the initial improvement, and include maintenance costs in the decision to expand.

Use the first month to establish a baseline and resolve access defects, the second to run a small content experiment, and the third to assess clicks, accuracy, useful actions, and costs. Low volume sites may require longer observation. Report inconclusive results honestly rather than selecting only prompts that improved.

8 Limitations and research priorities

Current evidence leaves several questions unresolved. How much does AI visibility influence later purchasing without a recorded referral? Which interventions retain their effect across model changes? Do small specialist publishers benefit from citation exposure, or does value concentrate among established platforms? What happens when many competing sites adopt the same optimization method?

Better studies would combine longitudinal user panels, contemporaneous capture of search interfaces, transparent sampling, and controlled interventions that measure downstream behavior. They should distinguish retrieval eligibility from answer prominence and compare economic outcomes across publisher business models. Null findings and failed replications are particularly valuable in a field where tools and consulting services create incentives to publish positive results.

This review is limited by its targeted search strategy, proprietary industry samples, and heterogeneous outcome definitions. It does not pool effects or estimate a worldwide trend. The research cutoff indicates when sources were checked; it does not imply that every dataset extends to October 2026. Platform controls and reporting capabilities should be rechecked before implementation.

9 Conclusion

Generative search changes the relationship between publishing useful information and receiving a visit. Available research supports pressure on some organic clicks, alongside evidence of changing referral quality and uneven opportunities for visibility. It does not establish the death of SEO, the permanent contraction of every informational category, or reliable financial compensation through citations.

A defensible strategy keeps technical SEO, useful content, and business measurement connected. It adds repeated tests of AI visibility and attribution accuracy, without treating either as the final objective. The organizations most able to adapt will understand where their own value is created and test whether a proposed change improves it.

References

Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Author manuscript, revised June 28, 2024. https://arxiv.org/abs/2311.09735

Chapekis, A., & Lieb, A. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Google. (n.d.). List of Google’s common crawlers. Google Crawling Infrastructure. Retrieved October 22, 2026, from https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers

Google Search Central. (n.d.-a). AI features and your website. Retrieved October 22, 2026, from https://developers.google.com/search/docs/appearance/ai-features

Google Search Central. (n.d.-b). Creating helpful, reliable, people-first content. Retrieved October 22, 2026, from https://developers.google.com/search/docs/fundamentals/creating-helpful-content

Google Search Central. (n.d.-c). Understand JavaScript SEO basics. Retrieved October 22, 2026, from https://developers.google.com/search/docs/crawling-indexing/javascript/javascript-seo-basics

Google Search Central. (n.d.-d). What is Googlebot. Retrieved October 22, 2026, from https://developers.google.com/search/docs/crawling-indexing/googlebot

Kim, S., Jeong, W., Kim, S., Lee, S., & Lee, D. (2026). SAGEO Arena: A realistic environment for evaluating search-augmented generative engine optimization (Version 2, August 7). arXiv. Listed as accepted at KDD 2026. https://doi.org/10.48550/arXiv.2602.12187

Law, R. (2026, February 4). Update: AI Overviews reduce clicks by 58%. Ahrefs. https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474. https://papers.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html

Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. Findings of the Association for Computational Linguistics: EMNLP 2023, 7001–7025. https://doi.org/10.18653/v1/2023.findings-emnlp.467

OpenAI. (n.d.). Overview of OpenAI crawlers. Retrieved October 22, 2026, from https://developers.openai.com/api/docs/bots

Pandya, V. (2026a, April 16). Adobe report: U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines. Adobe. https://business.adobe.com/blog/ai-traffic-surge-retail-sites-not-machine-readable

Pandya, V. (2026b, August 19). Adobe report: U.S. consumers are embracing LLMs to make travel plans, but many brands have AI visibility gaps. Adobe. https://business.adobe.com/blog/us-consumers-are-embracing-llms-to-make-travel-plans

Seer Interactive. (2026, April 24). AIO impact on Google CTR: 2026 update. https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update

Zhang, F., Cheng, Q., Wan, J., Singh, V., Rao, J., & Boakye, K. (2026). Generative engine optimization: A VLM and agent framework for Pinterest acquisition growth [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2602.02961