Here is the finding that should reorganise your 2026 link-building budget, and almost nobody is acting on it. When Ahrefs analysed 16.975 million cited URLs across ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews, the headline everyone repeated was that AI assistants prefer fresher content. True, but trivial. The buried, counter-intuitive result is this: the average age of a URL cited by an AI assistant is still 2.9 years, and Google’s own AI Overviews cite content that is 16 days older than the organic results sitting directly beneath them. (Ahrefs, 17M citation study)
If you read that study as “update your pages more often,” you read it backwards. The data does not reward churn. It rewards a very specific combination of signals that most link builders are not building for — and the gap between what the data says and what agencies are still selling is the single largest arbitrage opportunity in the discipline right now.
This article does something the original Ahrefs write-ups deliberately stop short of doing: it reverse-engineers the entire Ahrefs AI-citation research corpus — the 17M-citation freshness study, the 75,000-brand correlation studies, the AI Overview top-10 analysis, and the Q1 2026 AI Search Benchmark covering 13 studies, 146 million SERPs and 730,000 AI responses (Ahrefs / BusinessWire, May 2026) — into 12 patterns you can act on this week, plus a scoring formula you can run against any domain. If you are new to the underlying mechanics, start with our primer on what link building is in 2026 and come back.
The deliverable first: the Citation Probability Index (CPI)
Before the patterns, the tool — because the patterns only matter if you can measure where you stand. The Citation Probability Index (CPI) is a single 0–100 score that estimates how citation-ready a domain is for AI search, weighting each input by its measured correlation with AI visibility in the Ahrefs 75K-brand datasets. It is not a magic number. It is a way to stop guessing.
CPI = 100 × [ (0.74·Y) + (0.66·M) + (0.53·A) + (0.39·S) + (0.22·B) ] ÷ 2.54
Where each variable is a percentile (0 to 1) of your domain measured against your top 5–10 named competitors:
- Y — YouTube mention percentile (brand named in video titles, transcripts, descriptions)
- M — branded web-mention percentile (unlinked + linked mentions across the web)
- A — branded-anchor percentile (share of anchors containing your brand name)
- S — branded-search-volume percentile (total volume of queries containing your brand)
- B — referring-domain percentile (unique linking domains — the only classic link metric in the model)
The weights are not invented. They are the Spearman correlations Ahrefs published between each signal and AI visibility: YouTube mentions ~0.74, branded web mentions 0.66, branded anchors 0.53, branded search 0.39, and backlinks 0.22 (Ahrefs brand-visibility correlations; Ahrefs 75K-brand study). Dividing by their sum (2.54) normalises the score to 0–100. The structure itself teaches the lesson: four of the five inputs are off-page brand signals, and the lone link metric carries the least weight. That is the whole thesis of the 17M study in one equation.
How to read your CPI
| CPI band | What it means | Priority move |
| 0–25 | Below the visibility cliff. Effectively invisible to AI for category queries. | Fix the mention base before anything else. Micro-optimisation is wasted here. |
| 26–50 | Emerging. Occasional citations, highly inconsistent across re-runs. | Earned media + YouTube/podcast mentions to cross the threshold. |
| 51–75 | Competitive. Cited for some prompts, displaceable by rivals. | Referring-domain diversity and branded-anchor share. |
| 76–100 | Dominant. Cited reliably; the source competitors try to displace. | Defend with freshness on key assets; monitor for displacement. |
A worked CPI example
Suppose a mid-market SaaS brand scores, against its five closest competitors: YouTube mentions in the 30th percentile (Y=0.30), web mentions 45th (M=0.45), branded anchors 60th (A=0.60), branded search 50th (S=0.50), and referring domains 80th (B=0.80). Plug in: (0.74×0.30) + (0.66×0.45) + (0.53×0.60) + (0.39×0.50) + (0.22×0.80) = 0.222 + 0.297 + 0.318 + 0.195 + 0.176 = 1.208. Divide by 2.54 and multiply by 100 → CPI ≈ 48. The read is immediate: this brand has invested in links (80th percentile) but is starved of the signals that actually predict citation. Its largest weighted gap is YouTube mentions (0.74 weight × a 70-point percentile deficit). That, not more links, is the first move. The formula turns a vague “do more GEO” into a single ranked priority.
Run the CPI quarterly. The point is not the absolute number — it is the direction of travel and the gap to the competitor cited above you. Pull the inputs from any tool with an AI-visibility index; our guide to the best link-building tools breaks down which platforms expose mention and citation data natively.
What the 17M-citation study actually measured
The freshness study extracted ~17 million cited URLs from Ahrefs Brand Radar and calculated two ages for each: days since first publication, and days since last update. Then it compared AI citations against the plain organic SERP as a baseline. The aggregate result: AI assistants cite content that is 25.7% fresher by publication date (1,064 days vs 1,432 days) and 13.1% fresher by last-update date. (Ahrefs) The per-platform breakdown is where the actionable detail lives:
| AI surface | Avg days since published | Avg days since updated |
| Google AI Overviews (top 3) | 1,432 | 1,067 |
| Organic SERP (baseline) | 1,416 | 1,047 |
| Perplexity | 1,166 | 993 |
| Gemini | 1,118 | 831 |
| Copilot | 1,056 | 865 |
| ChatGPT (references) | 1,023 | 865 |
| ChatGPT (citations) | 958 | 989 |
Source: Ahrefs 17M citation study, July 2025. The spread between ChatGPT (958 days) and Google AIO (1,432 days) is 474 days — roughly 16 months. Treating “AI freshness” as one number erases the most useful finding in the dataset.
The freshness study is only the anchor. The companion Ahrefs studies measure why a URL gets selected in the first place — and that is where the 12 patterns come from. We have grouped them so you can act in order of leverage, not order of publication.
How AI retrieval actually selects a citation
To act on the patterns, you need the mechanism underneath them. A citation is the end of a three-step pipeline, and link builders can influence each step. First, the model makes a binary decision: answer from memory or search the web. Most users assume ChatGPT always searches — it does not. Writesonic’s March 2026 testing found GPT-5.4 Thinking skipped search on several of fifty prompts and, on one product query, cited seventeen sources from training data alone (Passionfruit, 2026). That is why brand presence in the training corpus — historic mentions, Wikipedia, durable forum threads — still matters even when live retrieval is off.
Second, if it does retrieve, the source overlap is platform-specific. ChatGPT matches Bing’s top-10 results roughly 87% of the time, which makes Bing optimisation directly relevant; Google AI Overviews pull about 76% of cited URLs from their own top-10 organic results; Perplexity leans heavily on Reddit (industry analysis). Third, from the retrieved set, the model decides what to extract — and this is where on-page structure earns or loses the citation.
The extraction-stage findings are remarkably consistent across independent studies, and they translate directly into editorial rules:
- The ski-ramp. Around 44.2% of LLM citations come from the first 30% of a page, and a further 31.1% from the middle (SparkToro / Kevin Indig analysis of 1.2M ChatGPT answers, via AI SEO statistics). Put the citable claim early — not after a 300-word preamble.
- Entity density. Heavily cited text averages ~20.6% entity density — three to four times normal English. Name the tools, people, companies and metrics explicitly rather than using pronouns and vague references.
- Readability inversion. Winning content averaged a Flesch-Kincaid grade of ~16 versus ~19.1 for lower performers — clearer, not dumber. Definite language and question-formatted headings help.
- Speed is a citation signal. SE Ranking found pages with first contentful paint under 0.4s averaged 6.7 citations, while pages slower than 1.13s dropped to 2.1. Technical performance is now a retrieval factor, not just a UX one.
None of this replaces off-page authority — it determines which of the already-authoritative pages gets extracted. The tooling to measure most of these signals is covered in our best link building tools round-up.
The 12 patterns, ranked by leverage
Pattern 1 — Freshness is a tiebreaker, not a ranking factor
The 25.7% freshness edge is real but bounded. The average cited page is still nearly three years old, and Google’s John Mueller has explicitly warned against changing publish dates without substantive content changes (Ahrefs). Action: earn fresh links and mentions pointing at evergreen assets rather than re-stamping dates. Freshness breaks ties between comparable sources; it does not manufacture authority where none exists. The practical version: find the three or four pages that already earn citations and direct newly earned links and mentions at those, rather than rewriting them. A link from a recently published article signals freshness to the target asset without disturbing the content already doing the work.
Pattern 2 — Platform-segment your freshness strategy
ChatGPT cites content up to 458 days newer than organic; Google AI Overviews cite content 16 days older. These are opposite incentives. Action: if your buyers live in ChatGPT and Perplexity, prioritise recency — timely commentary, refreshed data, newly earned coverage. If they live in Google AIO, depth and accumulated authority on long-lived URLs win. Stop running one undifferentiated “GEO” playbook across surfaces that reward the opposite behaviour.
Pattern 3 — Brand mentions beat backlinks roughly 3:1
Across 75,000 brands, branded web mentions correlated with AI Overview presence at 0.664, while backlinks managed only 0.218 (Ahrefs). That is close to a 3:1 advantage for being talked about over being linked to. Action: reallocate a slice of your link budget toward mention-generating activity — digital PR, expert commentary, podcast appearances — and stop measuring those campaigns on followed-link counts alone. The mention is the asset. The clearest test of whether your team has internalised this: ask what happens to a campaign that earns fifty unlinked mentions in relevant publications and zero followed links. Under the old scorecard it failed. Under the data, it may have done more for AI visibility than a month of link acquisition, and reclaiming a fraction of those mentions as links later is upside rather than the point.
Pattern 4 — YouTube mentions are the single strongest signal
In the December 2025 multi-platform follow-up, YouTube mentions (brand named in a title, transcript or description) correlated with AI visibility at ~0.737 across ChatGPT, AI Mode and AI Overviews — beating every other factor, including branded web mentions (Ahrefs). Google owns both YouTube and its AI surfaces, and they cite YouTube more than any other domain. Action: build a deliberate YouTube-mention surface — guest spots on niche channels, podcast video uploads, transcribed webinars — even if you never run a channel yourself. The mention does the work, not the subscriber count.
Pattern 5 — Branded anchors out-predict generic anchors
Branded anchor text correlated at 0.527 with AIO visibility and 0.628 with Google AI Mode (Ahrefs). For a decade, link builders suppressed branded anchors in favour of exact-match keywords. The AI era inverts that. Action: in outreach and guest contributions, ask for the brand name in the anchor. Our guest posting guide covers how to negotiate anchor text without tripping editorial resistance.
Pattern 6 — The visibility cliff: the bottom 50% is invisible
Ahrefs found that brands in the bottom half of web mentions are essentially invisible to AI Overviews — the relationship is not linear, it has a threshold (Ahrefs). Action: if your CPI is under 25, do not waste a sprint on schema or anchor tuning. Crossing the mention threshold is the only move that changes the outcome. This is also why displacement work has a floor — a point we cover in depth in our guide to AI citation recovery.
Pattern 7 — AI citations diverge sharply from rankings
Only about 12% of AI citations overlapped with Google’s top 10 in the August 2025 analysis, and roughly 80% of LLM citations did not rank in Google’s top 100 for the original query (Ahrefs). The retrieval evaluation and the ranking evaluation are measuring different things. Action: stop assuming a #1 ranking guarantees citation. Build retrieval-shaped assets (clear answers, structured data, named entities) in addition to ranking-shaped ones. For the broader strategic context, our link building strategies hub maps which tactics serve rankings versus retrieval.
Pattern 8 — Referring-domain diversity has a threshold effect
This is where backlinks still matter — but as diversity, not volume. SE Ranking’s data shows sites with up to 2,500 referring domains average 1.6–1.8 ChatGPT citations per category prompt, while sites above 350,000 referring domains average 8.4. The inflection point sits around 32,000 referring domains, where citation rates nearly double from 2.9 to 5.6. We unpack the displacement implications in our AI citation recovery playbook. Action: prioritise net-new unique linking domains over additional links from domains you already have. If you understand what backlinks are at a mechanical level, diversity-first acquisition is the obvious next step.
Pattern 9 — Citation order favours newest-first
Perplexity and ChatGPT both appear to order their in-text references from newer to older (Ahrefs). Position matters: the first cited source captures disproportionate attention and click-through. Action: time your highest-value earned media to land during active news or research cycles. Reactive, well-timed coverage can leapfrog older, higher-authority sources into the top citation slot — which is exactly the mechanic behind newsjacking for link building. The mechanism is worth naming: when a model orders references newest-first, the freshest credible source becomes the de facto primary citation, and primary citations capture the overwhelming majority of any click-through that survives the zero-click answer. A well-timed expert comment during a breaking story can therefore out-earn a year-old authoritative guide for that specific query window.
Pattern 10 — Distribution multiplies citations
Distributing the same content across a wide range of publications can increase AI citations by up to 325% compared with publishing only on your own site (Stacker, via AI SEO statistics roundup). Action: treat earned syndication as a citation lever, not just a traffic one. One study placed in fifteen outlets out-cites the same study published once. This is also why review-site and listicle placement has become so potent.
Pattern 11 — Format dictates retrieval: build citation-shaped assets
Ahrefs’ own analysis of which content gets cited points to a repeatable set of formats: original studies where you are the data source, clear glossary definitions, step-by-step how-tos, help documentation, and “best-of” listicles (Ahrefs, how to earn LLM citations). Each gives a retrieval system something specific to extract. Action: audit your linkable assets against these formats. The highest-yield format in 2026 is third-party listicle placement — we make the full case, with per-platform citation rates, in listicle placements: the new most powerful AI citation tactic.
Pattern 12 — Branded search is the downstream KPI
Branded search volume correlated with AI visibility at ~0.39 — weaker than mentions, but it is both a predictor and an outcome of citation success (Ahrefs). When AI cites you, more people search your name; that search then feeds back as a trust signal. Action: track branded search volume as the lagging indicator of a working GEO programme. It is harder to game than citation snapshots and closer to revenue. For the full measurement context, see our 2026 link building statistics hub.
The Monday-morning action table
Print this. Each row is a pattern, the metric that proves it, and the threshold that tells you whether to act. This is the operational core of the article — everything above is the evidence behind it.
| # | Pattern | Proof metric | Do this |
| 1 | Freshness = tiebreaker | Avg cited age 2.9 yrs; 25.7% fresher | Earn fresh links to evergreen pages; never date-stamp without edits |
| 2 | Segment by platform | ChatGPT −458d vs AIO +16d | Recency for ChatGPT/Perplexity; depth for Google AIO |
| 3 | Mentions > backlinks 3:1 | 0.664 vs 0.218 correlation | Move budget to PR/commentary; measure mentions not just links |
| 4 | YouTube mentions #1 | ~0.737 correlation | Build a YouTube/podcast mention surface this quarter |
| 5 | Branded anchors win | 0.527 (AIO) / 0.628 (AI Mode) | Request brand-name anchors in outreach |
| 6 | Visibility cliff | Bottom 50% mentions = invisible | If CPI < 25, fix mention base before optimising |
| 7 | Citations ≠ rankings | 12% top-10 overlap; 80% not in top 100 | Build retrieval-shaped assets, not just ranking ones |
| 8 | RD diversity threshold | ~32K RDs ≈ citations double (2.9→5.6) | Acquire net-new unique domains over repeat links |
| 9 | Newest-first ordering | Perplexity/ChatGPT order new→old | Time earned media to active news cycles |
| 10 | Distribution multiplies | Up to +325% citations | Syndicate studies across many outlets |
| 11 | Format dictates retrieval | Studies/how-tos/listicles cited most | Build citation-shaped assets; pursue listicle placements |
| 12 | Branded search = KPI | ~0.39 correlation, predictor + outcome | Track branded search as the lagging success metric |
Where each LLM pulls from: the 2026 source map
The patterns tell you what to build; this map tells you where to place it. Independent 2026 datasets converge on a clear hierarchy of cited domains, and it looks nothing like a traditional backlink target list.
- Reddit is the most-cited domain overall. Peec AI’s analysis of 30 million citations (March 2026) put Reddit first across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews combined. Perplexity in particular draws around 46.7% of its citations from Reddit. Forum participation has gone from a footnote to a frontline channel.
- LinkedIn is the #2 source for professional queries. Profound’s analysis of 1.4 million citations (Nov 2025–Feb 2026) found LinkedIn the most-cited domain for professional and B2B queries across the six major AI platforms. For B2B brands, an active LinkedIn presence is now a citation surface, not a vanity one.
- Review platforms punch above their authority. SE Ranking found domains with profiles on Trustpilot, G2, Capterra, Sitejabber and Yelp have roughly 3x higher odds of being chosen by ChatGPT than sites without them, and domains with heavy Quora/Reddit mention activity have ~4x higher odds. A complete review-site footprint is cheap, durable citation infrastructure.
- YouTube and Wikipedia anchor Google’s surfaces. Because Google owns AI Mode, AI Overviews and YouTube, those surfaces cite YouTube more than any other domain, and Wikipedia’s role in LLM knowledge graphs remains outsized. Both are long-horizon plays with high barriers — and high defensibility once won.
The strategic implication is uncomfortable for traditional outreach: a fresh, well-structured “best-of” listicle on a modest-authority site can earn more AI-citation visibility than twenty guest posts on higher-DR domains, because the listicle format aligns with how retrieval extracts content. We make the full, per-platform case in listicle placements: the new most powerful AI citation tactic. This source map is also the connective tissue for the rest of Phase 6’s operator-level GEO work — each of these domains deserves, and will get, its own dedicated playbook.
What the data shows vs what practitioners still believe
The reason these patterns are an arbitrage opportunity is that the prevailing beliefs in most agencies contradict the data. Four of the most expensive misconceptions:
Belief: “Update content constantly to win AI citations.”
The data: freshness gives a 25.7% edge between comparable sources, but the average cited page is 2.9 years old and date-only updates are explicitly discouraged. Constant updating has an opportunity cost; new, genuinely better content usually beats re-stamping. Freshness is a finishing move, not a foundation.
Belief: “More backlinks = more AI citations.”
The data: raw backlink volume correlates at just 0.218. What moves the needle is referring-domain diversity past a threshold, plus mentions and branded anchors. A thousand links from forty domains lose to four hundred links from four hundred domains. This does not mean links are dead — independent studies from Link.Build and Semrush both show topically relevant, high-quality links still lift citation odds — it means volume is the wrong target.
Belief: “Rank #1 and the AI will cite you.”
The data: 80% of AI citations do not rank in the top 100 for the query. Retrieval pulls from forums, listicles, YouTube and review sites that never ranked. Ranking helps, but it is neither necessary nor sufficient.
Belief: “Links don’t matter anymore for AI, only mentions.”
The data: the opposite over-correction. Backlinks still inherit the web’s authority graph that LLMs are trained on, and referring-domain diversity is Pattern 8 for a reason. Mentions lead, links support. A programme that abandons links to chase mentions hits the diversity ceiling and stalls.
Belief: “GEO is a separate discipline from link building.”
The data: the inputs are the same inputs link builders have always pulled — mentions, anchors, referring domains, earned media — just reweighted and measured against AI surfaces instead of (or alongside) the SERP. The 17M study does not announce a new profession; it re-prices the existing one. Treating GEO as a siloed team buying separate tools usually duplicates spend and fragments the very mention-and-link signals the data says should be coordinated.
A verifiable teardown: how one Ahrefs study became the cited source
The cleanest proof of Pattern 11 is the 17M study itself. Ahrefs published proprietary data nobody else had, in a format (an original study with extractable statistics) that retrieval systems love. The result is measurable and you can verify it yourself: that single blog post shows 199 referring websites on Ahrefs’ own backlink widget at time of writing, and it is now the source LLMs reach for when asked whether AI prefers fresh content (verify here). One asset, one proprietary dataset, a citation moat in a high-intent topic.
Here is the teardown method you can run on any competitor that out-cites you — reproducible, no guesswork:
- Identify a category prompt where a competitor is cited and you are not (e.g. “best [your category] tools”).
- Open the cited URL and classify its format against Pattern 11: study, glossary, how-to, listicle, review profile.
- Pull its referring-domain count and check diversity, not just volume (Pattern 8).
- Check publication and last-update dates against the platform’s freshness profile from the table above (Patterns 1–2).
- Search the brand on YouTube and count mention surfaces (Pattern 4).
- Score both domains with the CPI and read off the largest weighted gap — that is your first move.
Worked example by sector: in our teardown of the Indian SaaS market we found local brands crossing the citation threshold primarily through review-site profiles and regional PR rather than raw DR — the full breakdown is in link building in India and South Asia. In recruitment, we documented a site taking DR 0→17 in weeks purely from G2, Capterra and Crunchbase submissions before any outreach — see link building for recruitment and HR tech sites. Both confirm the same pattern: the citation foundation is built off-page, then links compound on top.
When NOT to act on these patterns
Honesty section, because every framework has a domain where it does more harm than good. Do not apply these patterns when:
- Your buyers are not in AI search yet. Some local-service and legacy-B2B niches still convert almost entirely through traditional SERPs. Google sends roughly 190x more traffic than ChatGPT today; over-rotating to GEO can starve the channel paying your bills.
- You are below the visibility cliff. If your CPI is under 25, Patterns 5, 9 and 11 are premature. Fix mentions first; optimisation on an invisible base yields nothing.
- You would treat correlation as causation. These are correlational studies on observational data. A 0.66 coefficient is a strong signal, not a guaranteed lever. Test on your own domain before betting a budget on it.
- You are measuring from a single snapshot. AI recommendations are highly inconsistent — there is well under a 1-in-100 chance two runs of the same prompt return the same brand list. Measure across many prompts and dates, or you will chase noise.
- Freshness churn would cannibalise evergreen equity. Re-writing a page that already earns citations to chase a recency signal can reset accumulated trust. If it is working, defend it; do not churn it.
A 90-day implementation sequence
The action table is prioritised by leverage; this is the same work sequenced by time, so a team can start Monday and have a measurable result inside a quarter.
Days 1–30: measure and clear the floor
Score your CPI against five named competitors. Run twenty category prompts across ChatGPT, Perplexity and Google AI Mode, three times each on different days, and log who is cited — this is your baseline, and the repetition is deliberate, because single snapshots are noise. If your CPI is under 25, your only job this month is the mention base: build out review-site profiles (G2, Capterra, Trustpilot), claim unlinked mentions, and line up two or three podcast or video appearances. Nothing else moves the number from here.
Days 31–60: build citation-shaped assets
Commission one piece of proprietary data — even a 200–300 respondent survey produces citable statistics — and structure it as a study with the headline finding in the first 30% of the page (the ski-ramp). Audit your three most important pages for entity density and first-contentful-paint, and fix the slow ones. Begin a deliberate listicle-placement campaign in your category. This is the phase that builds the assets retrieval systems extract.
Days 61–90: distribute, diversify, defend
Syndicate the study across as many relevant outlets as you can earn — distribution is the up-to-325% multiplier from Pattern 10. Push referring-domain diversity toward net-new unique domains rather than repeat links. Re-run the CPI and the twenty-prompt panel against the Day 1 baseline. The branded-search-volume line is your lagging confirmation: if it has started to lift, the programme is working and you defend the assets that earned it. If it has not, the teardown method tells you which weighted gap to attack next quarter.
The bottom line for 2026
Reverse-engineered, the Ahrefs corpus says something uncomfortable for a link-building industry built on followed-link counts: the strongest predictors of AI citation are brand signals, and the lone link metric that still matters is diversity, not volume. The agencies that win the next two years will sell mention generation, referring-domain diversity, citation-shaped assets and platform-segmented freshness — and they will measure success in branded search volume, not DR.
Start by scoring your domain with the CPI, find the largest weighted gap, and run the teardown method against whoever is cited above you. Then work the action table top-down. If you want the wider measurement picture, the 2026 link building statistics hub collects the benchmarks; if you are rebuilding a programme from scratch, start at the link building strategies hub and layer these 12 patterns on top.
One closing caution that the data earns the right to make: the coefficients in this article are snapshots of a system changing month to month. The direction of the findings — brand signals over raw links, diversity over volume, retrieval over ranking, distribution over self-publishing — has held across every major 2026 dataset and is unlikely to reverse. The exact numbers will drift. Build for the direction, measure your own domain against it, and re-score quarterly. The link builders who treat the Ahrefs corpus as a live instrument rather than a finished verdict are the ones who will still be cited in 2027.
Frequently asked questions
Do backlinks still matter for AI citations?
Yes, but as referring-domain diversity rather than raw volume. Backlinks correlate with AI visibility at only 0.218 versus 0.664 for brand mentions, yet diversity shows a threshold effect — citation rates roughly double around 32,000 referring domains. Links inherit the authority graph LLMs train on; they support the mention signals that lead. The mistake is chasing link count instead of unique-domain breadth.
How often should I update content for AI search?
Less often than the freshness headline implies. The 25.7% freshness edge is a tiebreaker between comparable sources, and the average cited page is still 2.9 years old. Update when you have a substantive change — new data, corrected facts, a better answer — not on a calendar. Date-only updates are explicitly discouraged and can reset accumulated trust on pages that already earn citations.
Which AI platform should I optimise for first?
Whichever your buyers actually use, then segment. ChatGPT and Perplexity reward recency and lean on Reddit and Bing’s index; Google AI Overviews reward depth, accumulated authority and their own top-10 organic results. If you do not know where your buyers are, start by measuring referral traffic by source and follow the data rather than the hype.
How do I measure whether I’m getting cited?
Use three methods together: manual prompt testing across platforms run repeatedly on different days; GA4 segments for AI referrers; and a dedicated AI-visibility tool that exposes citation and mention data. No single method is sufficient because AI answers are inconsistent across runs. The lagging KPI to trust is branded search volume — harder to game and closer to revenue than any citation snapshot.
Is GEO replacing link building?
No — it is re-pricing it. The strongest predictors of citation are the same earned signals link builders have always pursued: mentions, branded anchors, diverse referring domains and distributed earned media. The weights changed and the measurement surface expanded to include AI answers. Teams that coordinate link building and GEO under one strategy outperform those that silo them.
