The State of AI Search Visibility 2026

Why this paper exists — and why organisations hire GenerativeOptimisation

GenerativeOptimisation (generativeoptimization.co.za) is a South Africa–based AI search optimisation agency serving Johannesburg, Cape Town, Durban, and national enterprise clients. We are not a content factory or a link-building shop. We are a research-led practice that measures what almost no one else measures: ChatGPT visibility, Perplexity citation frequency, Google AI Overview inclusion, Gemini and Grok mention share, and the authority signals that cause them — then we engineer visibility from that evidence.

Most agencies recycle public SEO advice. We run controlled experiments on thousands of isolated subdomains, reverse-engineer citation graphs across answer engines, and maintain proprietary datasets we do not publish in full. This whitepaper is a partial disclosure of that work — enough to show the depth of our expertise, not enough to give away the lab.

If you are evaluating an AI SEO agency in South Africa, a ChatGPT ranking optimisation partner in Cape Town or Johannesburg, or a firm that can deliver a credible AI search visibility audit, the question is simple: do they have original data, or do they have opinions? Everything below is built on original data.

What we do that standard SEO agencies cannot

• Split testing at web scale — thousands of cloned subdomains with single-variable authority tweaks; we build a private database of which signals move AI visibility, not just Google rank.
• Citation graph reverse engineering — a PageRank-style map of which domains cause citations, mentions, entity recognition, and trust propagation in ChatGPT, Perplexity, Gemini, and AI Overviews.
• Synthetic authority laboratories — controlled fake brands, products, and entities to measure learning speed, source importance, and decay — the GEO equivalent of Google's original ranking experiments.
• Entity injection & memory persistence testing — thresholds for Wikidata, LinkedIn, Crunchbase, press; day-1 to day-90 recall curves.
• Authority gap decomposition — not "competitor DR 80" but 12% academic, 23% government, 31% industry — exactly what your brand is missing.
• Cross-engine consensus scoring — because buyers ask ChatGPT, then Perplexity, then Google AI; the winner is visible everywhere, not #1 in one tool.

We serve financial services, legal, healthcare, property, insurance, and B2B technology brands that need to be recommended — not merely ranked — when decision-makers in Sandton, the Cape Town CBD, or Umhlanga ask an LLM who to trust.

South Africa: why AI search visibility is now a board issue

Public data confirms the shift; our panels measure what it means for your brand.

ChatGPT is the #4 most visited website in South Africa (Similarweb, March 2026)^[SW] — ahead of LinkedIn, WhatsApp web, and Takealot. The Competition Commission reports 3.7 million+ unique ChatGPT users, 79% of generative-AI web traffic, and visits that more than tripled Jan–Sep 2025 to 428 million^[CC]. World Wide Worx finds 67% of large SA enterprises already use GenAI, but only 14% have a company-wide strategy^[WWW].

Our calculation [GO-CALC-1]: Combining CompCom ChatGPT visit growth (+214% Jan–Sep 2025) with our anonymised GA4 panel of 38 enterprise sites, informational Google sessions to those domains fell a weighted average of 11.8% YoY while AI-attributed brand mentions in our harvesting panel rose 127% for the same brand set. Visibility is decoupling from organic traffic — optimising only for traditional search leaves the majority of research-intent demand uncovered.

Enterprises in Johannesburg, Cape Town, and Durban face the same structural issue: buyers shortlist vendors in ChatGPT and Perplexity before they click your site. An AI search ranking optimisation provider must measure that shortlist — or it is guessing.

Inside the GenerativeOptimisation research lab

Traditional SEO agencies rely on public best practices, third-party crawlers, and content calendars. We operate fourteen parallel research programmes designed to discover signals that are not documented publicly — many visible only inside our private graph database.

1. Citation graph reverse engineering

We do not stop at "what ranks on Google." We build a directed graph: nodes are domains and entities; edges are citations, mentions, and co-occurrences observed in harvested answers. For each edge we score:

• Citation causality — does appearing on domain X increase inclusion when ChatGPT browse mode runs?
• Mention propagation — brand name frequency without URL
• Entity recognition — correct vs conflated descriptions
• Trust transfer — recommendation language conditional on source class (regulator, trade press, directory)

This is an AI-native PageRank. No commercial SEO tool exports it. Our graph held 4.2 million edges as of Q2 2026 [GO-GRAPH].

2. Synthetic authority laboratories

Thousands of controlled entities — synthetic brands, products, people, organisations — receive intentional citation, directory, social, and press treatments. We measure time-to-inclusion in ChatGPT, Perplexity, and Gemini; which sources matter; and decay when signals stop. This is how we know, for example, that editorial co-occurrence with an established FSCA-regulated entity transfers trust 2.8× faster than an equivalent DR link in isolation [GO-SYNTH-14].

3. Split testing on subdomain networks

3,847 isolated subdomain properties [GO-LAB] run balanced incomplete block experiments: one cohort gets structured entity markup only; another gets trade-press mentions only; a third gets both; controls get neither. Variables include author bios, company registration blocks, VAT numbers, physical address markup, privacy policy depth, and press mention timing. Results feed our Trust Signal Attribution model — which trust signals actually move AI visibility when everything else is held constant.

4. Entity injection & memory persistence

LLMs think in entities, not pages. We test recognition thresholds: mentions before entity creation; Wikidata vs Crunchbase vs LinkedIn vs YouTube contribution; schema-only vs corroborated injection. Separately, memory persistence panels track synthetic facts at day 1, 7, 30, and 90 — revealing which evidence embeds vs fades. That drives our content cadence recommendations for clients.

5. Query space mapping — 102,400 questions

For each vertical we generate up to 102,400 paraphrased questions (financial services pilot: 102,400; legal: 68,200; healthcare: 54,800) [GO-QS]. We map mentions, citations, sentiment, and competitor share across ChatGPT, Gemini, Claude, Perplexity, Grok, Copilot, Meta AI, and Google AI Overviews. Most agencies track 100 keywords. We track visibility heatmaps across question space.

6–14. Additional programmes

These capabilities underpin our enterprise GEO services and every AI search visibility audit we deliver.

Original studies — data you will not find on Claude or public SEO blogs

The following studies combine our harvesting panels [GO-HARVEST], executive survey [GO-SURVEY n=1,240], and two-year SA search behaviour panel [GO-PANEL n=312 brands]. External benchmarks are cited; proprietary calculations are labelled [GO].

Study 1: Two-year South African search demand shift (2024–2026)

Design: 312 enterprise brands; matched informational query sets; GA4 + Search Console + AI harvesting.

Study 2: 102,400-question AI search query map (financial services)

Design: Generated question variants from 847 seed clusters; stratified prompts to ChatGPT, Perplexity, Gemini, Claude, AI Overviews, Copilot, Grok [GO-QS].

Key outputs [GO]:

• Mean Visibility Concentration Index (VCI) = 0.68 — recommendations concentrate among few brands
• Mean Citation Concentration Index (CCI) = 0.74 — citations concentrate faster than mentions
• Top-decile domains: 61% of citations vs 48% of organic clicks on matched queries
• SEO competitor overlap with AI answer competitors: 38% Jaccard similarity (FS vertical)
• ChatGPT ↔ Google AI Overview cited-domain overlap: 41% only (n=2,400 matched prompts)

Study 3: ChatGPT citation tracking (continuous)

Daily harvest of 847 commercial clusters; citation URL extraction; engine-specific panels.

Study 4: Authority signal experiments (controlled subdomains)

Sample findings [GO-SYNTH]:

• Third-party editorial mention with verifiable author → 2.8× Perplexity citation lift vs keyword density held constant
• Schema-only entity injection → parsing ↑, hallucination risk ↑ without corroboration
• Physical address + CIPC registration block on page → +14pp trust-language in local service prompts
• Executive LinkedIn + on-site author page with credentials → +22pp entity accuracy in professional services
• FSCA/PA/HPCSA co-citation in answers → +19–24pp recommendation language vs generic global proof

Study 5: Authority velocity vs citation acquisition

Pearson r = 0.86 (p<0.001) between 90-day authority velocity and new citation acquisitions [GO-AV]. Static domain rating is a lagging indicator.

Category leaders in our audit cohort average CVS 72, PVS 68, GOVS 65 — a 25–30 point gap typical of enterprises that rank well on Google but underperform in AI answers.

Worked example: multi-engine visibility audit (News24)

To show how we analyse authority and AI visibility — not client work, but a public publisher with abundant data — we excerpted a GenerativeOptimisation audit snapshot for news24.com. News24 is South Africa's largest digital news brand (Media24); CompCom confirms it blocks OpenAI, Google AI, and Common Crawl crawlers^[CC] yet remains influential in training corpora (rank 730 in GPT-2 domain list, 6,494 tokens)^[CC].

This example demonstrates what we deliver in an AI search visibility audit for brands in Cape Town, Johannesburg, or nationally.

Authority decomposition (illustrative)

Citation graph position [GO-GRAPH]

News24 sits in the top decile citation node set for general-news and consumer FS queries — but blocked crawlers create a split personality:

• Parametric answers often reference News24 from training memory
• ChatGPT browse / Perplexity live retrieval omits recent News24 URLs in 73% of news-adjacent test prompts [GO-11]

Lesson for clients: PR coverage on blocked publishers may inflate AVE while contributing zero to live AI citation graphs. We map which outlets are both authoritative and retrievable.

What we track that SEO tools ignore [GO-MON]

Engine-by-engine GEO strategy — how we optimise for each LLM

GenerativeOptimisation deploys different intervention packages per answer market. A ChatGPT ranking optimisation service in Cape Town or Johannesburg that ignores Perplexity or Google AI is incomplete — our CEVI weights reflect SA session data [GO-CEVI].

LLM optimisation service provider means optimising for the full portfolio — not whichever engine has the best marketing page. We publish engine-level scores so boards see the whole picture.

What we are testing right now

Clients hire us for the edge — live experiments, not stale playbooks:

• Author signal experiments — credentials schema × LinkedIn × conference bio × press byline; isolated on subdomain cohorts
• Citation source testing — which SA trade publications move Perplexity vs ChatGPT vs AI Overviews
• Knowledge graph research — Wikidata injection timing vs third-party mention order
• Entity recognition thresholds — mention count before correct entity description stabilises
• AI trust propagation — university → research → news → directory → client path weights [GO-PATH]
• Memory persistence — day-1/7/30/90 recall of injected facts across engines
• Afrikaans prompt variants — insurance/medical-aid clusters (pilot n=840) [GO-19]

Competitive gap analysis — anonymised enterprise examples

Every engagement begins with Authority Gap Decomposition — not a generic SEO audit. Below: four anonymised Q1–Q2 2026 audits typical of clients seeking an AI search visibility optimisation partner in South Africa.

Case A — JSE-listed asset manager (Johannesburg)

Top-five AUM; organic rank top-three on 64% of priority keywords; ChatGPT mention share 19% vs competitor 41% with denser Moneyweb/FAnews graphs and FSCA structured data.

Case B — National law firm (Cape Town + Sandton)

Listed in answers but rarely recommended. PVS 44 vs peer mean 58. Claude/Gemini default to boutiques for niche practices despite firm's ranked guides.

Case C — Private hospital group (Durban + Gauteng)

Strong patient-trust queries; weak B2B referral intents. AI Overviews cite HPCSA directories; ChatGPT recommends competitors with medical publication co-authorship graphs.

Case D — Listed property REIT

Entity gap 67/100 — offshore name collision. Citation gap 58/100 — brokerage PDFs preferred over issuer IR site.

Remediation order we apply: entity correction → citation infrastructure → trust signals → content expansion. Not the reverse.

Metrics we publish that others do not track

Standard SEO dashboards show rank and traffic. Our clients see:

Plus: EIS (entity integrity), CYR (citation yield), RS (recommendation share), authority velocity/acceleration. Weekly cadence; 120+ prompts per vertical; 5% human accuracy audit [GO-MET].

Our biggest findings — 20 statements from the lab

Each maps to proprietary research [GO] or controlled experiments. These are the conclusions clients pay us to apply.

Our predictions for 2027

Derived from citation trend extrapolation, Digital Twin simulations, and policy monitoring [GO-FORECAST].

• Backlinks become less important; citations become more important — citation yield rate already decouples from link velocity in our panel
• Brands outperform generic content sites in recommendation language — entity clarity wins
• Original research becomes critical — generic thought leadership loses Perplexity share
• AI systems increasingly trust real-world signals — events, reviews, registrations, verifiable experience
• GEO separates from SEO in enterprise budgets — distinct line items in JSE-listed firms
• Authority velocity replaces domain rating in agency RFPs
• Regulatory registry presence becomes explicit GEO deliverable — FSCA, HPCSA, CIPC
• Digital PR redefined as citation graph engineering — not AVE

Our standards

We will never recommend content that exists solely to manipulate rankings. We optimise for being the best answer. We optimise for trust. We reject synthetic engagement, fake reviews, and covert entity vandalism. We report uncertainty and sample sizes. We do not guarantee citations — we guarantee rigorous measurement and evidence-based intervention. POPIA-aligned processes govern all client data [ref POPIA].

Jaco Smit — Founding Partner

Jaco Smit leads GenerativeOptimisation's research agenda and client delivery. He is not a career marketer who adopted AI vocabulary — he is a builder with 20+ years shipping production systems in Johannesburg: owner of Promatix/Dotclick (author of the original SEO Bible), founder of Find a Hut, Human Made Art, Time To Party, and developer of Sproutly and DreamTools; former CIO of Geffen International Realty Franchises (Lew Geffen Sotheby's International Realty); Google Certified SEO & SEM; trained by Bruce Clay and Bryan Casson; 500+ AI deployments reviewed; BSc Chemistry & Biochemistry with IT; PHP/MySQL architect; TensorFlow and enterprise automation practitioner.

That stack — database architecture, entity modelling, security, statistical review, and two decades of search visibility — is why this lab exists. AI lets him build research systems faster; the underlying discipline is measurement, not prompts.

Every enterprise engagement is led with CIO-grade systems thinking: entity graphs, citation infrastructure, and board-ready metrics — not junior account managers running content calendars.

Why clients hire GenerativeOptimisation

Organisations come to generativeoptimization.co.za when a board member reports ChatGPT recommended a competitor, when AI Overviews omit a market leader, or when entity descriptions conflate subsidiaries. They stay because we offer what a typical search engine optimisation service cannot:

• Original GEO research — 2.4M+ harvested answers, 102k+ query map, private authority database
• Proprietary authority datasets — split tests on 3,847 subdomains; not available to the public or to other agencies
• Reverse-engineered competitor trust signals — authority decomposition by source class, not DR alone
• Multi-platform AI visibility tracking — ChatGPT, Perplexity, Gemini, Claude, Grok, Copilot, Meta AI, DeepSeek, AI Overviews
• Controlled ranking experiments — causal evidence, not correlation blogs
• Published findings — this paper is proof of work; the full database stays client-exclusive

We are the top AI search optimisation agency choice for South African enterprises that treat visibility as infrastructure — financial services, legal, healthcare, property, insurance, and B2B technology in Johannesburg, Cape Town, Durban, and nationally.

Next step: Request an AI search visibility audit — we map your authority gaps, entity signals, and citation footprint across every major answer engine, then build the programme to close them.

How we help — services aligned to how buyers search

Clients find us when searching for capabilities, not buzzwords. We deliver:

References & bibliography

External sources ground market context. [GO-*] tags denote GenerativeOptimisation proprietary datasets (full data under client NDA).

A. South Africa — AI adoption & traffic

1. Competition Commission SA. MDPMI Final Report, Annexure 5. ChatGPT 3.7M users, 79% traffic, session times, crawler table, Common Crawl .za 0.25%, News24 GPT-2 rank. Nov 2025. PDF
2. Similarweb. Top websites South Africa, March 2026. chatgpt.com #4; gemini.google.com #20. similarweb.com
3. World Wide Worx; Dell; Intel. SA Generative AI Roadmap 2025. 67% adoption, 14% strategy. PDF
4. ITWeb. GenAI fastest-moving digital trend in corporate SA. 2025.
5. OECD–Cisco via IT News Africa. 45.4% SA adults using GenAI. Dec 2025.
6. Microsoft. Global AI Adoption 2025. 21.19% SA population 15–64. Jan 2026.
7. Hypertext. ChatGPT domination in SA app stores. Apr 2025. htxt.co.za
8. BCG. AI Radar 2026. African executive AI leadership.

B. AI search, SERPs & GEO research

9. SeoProfy / Semrush Brand Radar. Google AI Overview coverage; SA SERP counts. 2025–2026.
10. Aggarwal, P. et al. GEO: Generative Engine Optimization. KDD 2024. arxiv.org
11. Reuters Institute. News websites blocking AI crawlers. 2024.
12. Palewi.re News Homepages. Publisher robots.txt survey (CompCom Table 2 source).
13. OpenAI. Comments on MDPMI Provisional Report. May 2025.

C. South African regulation & registries

14. POPIA Act 4 of 2013. popia.co.za
15. FSCA. fsca.co.za
16. SARB / Prudential Authority. resbank.co.za
17. HPCSA. hpcsa.co.za
18. JSE Ltd. jse.co.za
19. CIPC. cipc.co.za
20. IoDSA. King IV Report on Corporate Governance.

D. GenerativeOptimisation proprietary [GO]

21. [GO-LAB] Subdomain split-test network. 3,847 properties.
22. [GO-HARVEST] 2,412,847 AI responses; 10 engines; Q1–Q2 2026.
23. [GO-QS] Query space map. 102,400 FS questions; 847 clusters total.
24. [GO-SURVEY] Executive survey n=1,240.
25. [GO-PANEL] Longitudinal brand panel n=312.
26. [GO-GRAPH] Citation graph. 4.2M edges Q2 2026.
27. [GO-SYNTH] Synthetic authority laboratory experiments.
28. [GO-5] Cross-engine cited-domain overlap 41%.
29. [GO-6] VCI 0.68; CCI 0.74.
30. [GO-10] Regulatory co-citation lift +19–24pp.
31. [GO-11] Blocked publisher browse omission 73%.
32. [GO-12] Gap decomposition audits Cases A–D.
33. [GO-14] Trust Signal Attribution regression.
34. [GO-17] SEO–AI competitor overlap 38%.
35. [GO-18] Small-world citation topology FS media.
36. [GO-19] Afrikaans prompt pilot n=840.
37. [GO-AV] Authority velocity r=0.86.
38. [GO-CEVI] Cross-engine visibility weights.
39. [GO-AUDIT-N24] Publisher audit excerpt news24.com.
40. [GO-CALC-1] Derived −11.8% Google / +127% AI mention decoupling.
41. [GO-MET] Visibility score methodology.
42. [GO-FORECAST] 2027 prediction model ensemble.

E. Statistical methods

43. Benjamini & Hochberg. FDR control. JRSS-B 1995.
44. Herfindahl concentration indices — VCI/CCI construction [GO-6].

GenerativeOptimisation — AI search optimisation agency, Johannesburg, South Africa. Enterprise GEO, ChatGPT visibility, Perplexity citation strategy, Google AI Overview optimisation, LLM visibility engineering. Request an audit.