Tag: agencies

If you’ve been watching your AI visibility tracker the past six months, you’ve probably noticed the same thing I have: brands that get cited beside category leaders end up cited more — for queries they didn’t even target. That’s not luck. That’s the entity graph doing its job, and most operators are still pricing the implication in too slowly.

The old game was the link graph: who points to you, with what anchor text. The new game is the entity graph: who you’re mentioned beside, in what kind of source, in what shape of sentence. ChatGPT, Perplexity, Gemini, and Google AI Overviews all build internal representations of your brand from the company you keep across the training and retrieval corpus. Showing up alone on your own marketing pages doesn’t move the model’s picture of you. Showing up in a paragraph that already names two trusted category players does.

How the co-mention signal actually works

Embeddings don’t read your homepage and decide what you do. They look at every context you appear in across the corpus and cluster you with whatever you keep showing up next to. Two consequences follow.

First, when someone prompts ChatGPT or Perplexity with “what are the alternatives to [category leader],” the model surfaces brands whose embedding sits close to that leader’s — meaning brands consistently named in the same paragraph, the same comparison table, the same roundup post. Not the brands with the most backlinks. The brands with the most adjacency.

Second, the citation-share data lines up structurally. Roughly 47.9% of ChatGPT’s citations come from Wikipedia, and a comparable share comes from directory and listing sites — the exact surfaces where competitor sets get bundled into single paragraphs. When an LLM cites a “best CRM” listicle, it rarely cites the listicle’s pick #1 alone. It surfaces the whole comparison set. If you’re in the set, you’re in the answer.

The practitioner mistake

Most teams treat third-party placement as a generic citation-harvest play: “get on more roundups, get cited more.” That’s only half the lever. The other half is who you sit next to on those roundups. A “best SaaS tools 2026” mention beside fifteen brands nobody recognizes teaches the model nothing useful about you. A “alternatives to [category leader]” placement that names you, the leader, and two other recognized players teaches the model exactly what you are, instantly.

So the audit question isn’t “where can I get listed.” It’s “which adjacencies will define me in two years.”

What to do this week

Run the adjacency audit. Pick five prompts an ideal customer would actually type into ChatGPT, Perplexity, Gemini, and Google AI Overviews. Examples: “best [category] tools for [your ICP],” “alternatives to [category leader],” “[category leader] vs [smaller competitor],” “top [category] companies in 2026,” “cheaper alternatives to [category leader].” Note every brand that surfaces beside you — and every brand that surfaces instead of you. That’s your real positioning, not the one in your deck.

Target the right third-party placements. From that audit, identify the specific roundups, comparison pages, and review sites where your target adjacents already appear. Pitch yourself onto those. A spot on a list that bundles you with the right two competitors is worth ten spots on lists nobody mines.

Earn co-mention through original work. Get quoted in articles that name your target adjacents. Co-author a piece with a credible analyst. Take the podcast guest slot on a show that just had the category leader on. Each of those creates a co-occurrence record that the next training pass — and every retrieval-time index — absorbs.

Shape your own adjacency signal. Most brands write about themselves in isolation. Add an explicit “how we compare to [adjacent player]” section on your comparison page. Mention the two competitors you want to be associated with in your case-study language and your FAQ answers. You can move the needle on your own pages — competitors won’t object, and the entity graph will absorb the pattern.

Track citation share, not just citations. A citation alone is a vanity metric in the AI era. The number that matters is the percentage of “alternatives to X” answers that surface you in the set. If that share moves from 0% to 30% over a quarter, you’ve done the work. If it stays at 0% while your raw citation count climbs, you’re getting cited in the wrong neighborhoods.

Agencies: if your clients are starting to ask about AI SEO and you don’t have anyone in-house, Paris Roussos handles the work white-label — flat-rate, $500–$1,500/mo per end client, you keep the relationship. Email parisroussos@gmail.com for a sample audit.

The brands that win AI search over the next two years aren’t going to be the loudest — they’ll be the ones the model can place precisely on the map.

There’s a difference between a page that’s old and a page that looks old to a machine. Most operators have heard that AI search rewards fresh content. What almost nobody has fixed is the quieter problem underneath it: AI engines can’t apply a freshness preference to a page if they can’t tell when it was written. And a startling number of pages — including good ones — give the machine nothing to work with.

This isn’t the “refresh your content every two months” advice. That’s about changing the page. This is about signaling the change. You can refresh a page religiously and still lose the citation because the recency signal never makes it into a form an engine can read.

Why an engine needs a date before it can trust you

Freshness is one of the heaviest weights in AI retrieval right now. Roughly 65% of AI bot crawl activity targets content published within the past year, and pages updated within the past two months earn about 28% more citations than older ones. That’s the lever everyone’s chasing.

But here’s the mechanic: that lever only fires when the engine can place your page on a timeline. It establishes a date from three things — the `datePublished` and `dateModified` fields in your Article schema, a visible dateline near the headline, and contextual clues in the body copy (“as of 2024,” “last year’s data”). When all three are missing or contradictory, the engine doesn’t assume your page is fresh. It assumes it’s undateable, and undateable content gets deprioritized for anything time-sensitive — which, in 2026, is most commercial queries.

Worse is the page that’s actively misdated. A huge share of WordPress and CMS-built pages carry a `datePublished` from years ago and have never emitted a `dateModified` at all. You updated the body three times; the schema still says 2021. The engine reads the only date it was given and files you under stale. You did the work and got penalized for it because the signal lied.

The three places your date has to live

A date that exists in only one place is a date the engine might miss. It has to be consistent across all three surfaces, or the contradiction itself becomes a trust problem.

Machine-readable: your Article (or BlogPosting) schema needs both `datePublished` and `dateModified` as valid ISO timestamps. Not one. Both. The gap between them is what tells an engine the page is maintained rather than abandoned.

Visible: a human-readable “Published / Updated” line near the H1. AI engines parse rendered text, and a visible dateline corroborates the schema. Two surfaces agreeing is far stronger than one surface alone.

Contextual: the body copy itself. A page that says “the current state of AI Overviews” with no year reads as timeless; a page that says “as of May 2026” hands the engine a hard anchor and a recency claim in one phrase.

What to do this week

1. Audit your top 20 pages. Open each in a structured-data tester and check whether `dateModified` exists and is recent. You’ll likely find a third of them have no `dateModified` and a stale `datePublished`. That’s your fix list.

2. Add both date fields to your Article schema template, and wire `dateModified` to update automatically whenever the page is edited — not manually, because manual always drifts.

3. Surface a visible “Updated” date near every headline. If your theme hides it, unhide it. The corroboration is worth more than the cleaner design.

4. When you refresh a page, change something real. Engines and Google both detect date-only edits where the content didn’t move, and a bumped `dateModified` with no substantive change can read as manipulation. Update the date because you updated the page, never the reverse.

5. Kill orphan year references. Search your body copy for naked years and either remove them or make them deliberate. “In 2023, marketers…” on an evergreen page is a self-inflicted aging signal.

Paris Roussos has been doing SEO since 1996 (co-founded a Forbes Best of the Web–winning site back in the day) and now runs a white-label AI SEO practice for agencies and brands — flat-rate, $500–$1,500/mo per client. If your top-of-funnel traffic is leaking into ChatGPT and Perplexity and you want it back, email parisroussos@gmail.com.

Freshness isn’t only about doing the work — it’s about making sure the machine can see that you did.

Open ChatGPT, Perplexity, or Google’s AI Overview and ask “what’s the best CRM for a small agency” or “Asana vs Monday for a five-person team.” Watch what the answer is built from. It is almost never a vendor’s homepage and almost never a generic blog post. It is a comparison page — a head-to-head, a “best of” roundup, a category breakdown. If you sell anything that gets compared, the comparison page is now the single highest-leverage asset you can publish for AI visibility. Most operators are still treating it as an afterthought.

Here is why this happens, and what to do about it this week.

The mechanic: comparison content matches the shape of the prompt

LLMs answer evaluative questions far more often than factual ones. “Best,” “vs,” “alternative to,” “is X worth it” — these dominate the query mix because that is what people actually want from an AI assistant: a recommendation, not a definition. When the model retrieves sources to ground that answer, it reaches for content whose structure already mirrors the question. A comparison page does this natively. It names the contenders, lists evaluation criteria, weighs trade-offs, and lands a verdict. The model can lift a row, a criterion, or a one-line judgment and drop it straight into the answer with minimal rewriting.

A homepage cannot do that. Marketing pages assert that you are the best without showing the comparison work, so the model has nothing structured to extract. A standard blog post buries the comparison inside prose, so retrieval has to guess at the relevant passage. The comparison page front-loads the exact answer unit the model needs — and front-loading matters: roughly 44% of LLM citations come from the first 30% of a page. A page organized around the decision puts that decision near the top by design.

There is a second reason comparison pages punch above their weight: they read as neutral. AI systems lean toward sources that appear to weigh options rather than sell one. A page that honestly says “competitor X is better for enterprise, we are better for small teams” gets treated as an assessment, not a pitch — and assessments get cited. Pages that contain explicit criteria, named trade-offs, and concrete figures get pulled more often than pages that just praise. The same dynamic shows up in the broader citation data: pages dense with statistics and direct comparisons earn a 20%-plus visibility lift over thin, claim-only content.

The uncomfortable part for vendors: the comparison page that gets cited does not have to be yours. Third-party roundups, Reddit threads, and review-site category pages are filling that slot right now. If you are not publishing your own honest comparison content, you are conceding the most-cited surface in your category to someone whose verdict you do not control.

What to do this week

Build the three comparison pages your buyers actually search. Not twenty — three. Your product vs your top-named competitor, your product vs the second one, and a “best [category] for [your ICP]” roundup that includes you honestly among four or five options. These three pages map to the three prompts your buyers are already typing into AI assistants.

Lead every comparison with a structured answer unit. Open with a real table — criteria down the side, options across the top — and a 40-to-60-word verdict directly beneath it. That table and verdict are the answer block a model will quote. Bury the comparison under 600 words of preamble and retrieval will skip it.

Be honest enough to be useful. Name at least one scenario where a competitor wins. A page that only ever concludes “we win” reads as marketing and gets discounted. A page that says “choose them if you need X, choose us if you need Y” reads as a decision aid — and decision aids get cited and, frankly, convert better too.

Add concrete figures and FAQPage schema. Replace “more affordable” with the actual price, “faster setup” with the actual number of days. Then mark the page up with FAQPage and, where relevant, structured comparison data so the criteria are machine-legible. The schema will not earn the citation by itself, but it removes ambiguity about what each row means.

Agencies: if your clients are starting to ask about AI SEO and you don’t have anyone in-house, Paris Roussos handles the work white-label — flat-rate, $500–$1,500/mo per end client, you keep the relationship. Email parisroussos@gmail.com for a sample audit.

The brands winning AI search in 2026 are not the ones shouting that they are best — they are the ones publishing the comparison the buyer was going to run anyway.

The fastest way to get cited by ChatGPT, Perplexity, Gemini and Google’s AI Overviews in 2026 is not better copywriting. It is not a smarter schema kit. It is owning a number that nobody else has, and putting it on a page LLMs can read.

Every operator I talk to is grinding on the same playbook — refresh the post, tighten the H2s, ship an FAQ block, get the schema right. All of that helps. None of it gets you cited the way one defensible original statistic does. Citation engineering moves you up the rank inside a contested topic. Original data takes you out of the contest entirely.

Why LLMs reach for first-party numbers

Generative answer engines are answer-makers, not opinion-makers. When a model produces a response that says “X grew by 37% in Q1,” it needs a source it can defensibly point to. There are only so many sources that satisfy that pattern: government data, big-platform telemetry, analyst reports, and your blog post — if your blog post is the only thing on the open web that contains that exact number.

This is why the same studies keep showing that pages built around statistics and quotations get cited at materially higher rates than pages built around generic argument. We’ve already covered the +22% lift for stat-heavy pages and the +37% lift for quote-heavy pages on this blog. Both of those uplifts apply to secondhand stats and quotes — facts you pulled from someone else. When the number is yours and lives nowhere else, the citation behavior compounds. The model has no alternative source to fall back to. You become the alternative source.

Look at who dominates AI citations today. Reddit, Wikipedia, Stack Overflow, Statista, Gartner, McKinsey, Pew, BLS, government datasets. Notice the pattern: every one of those is either user-generated content the model has no other route to, or original research the model is forced to attribute. Almost none of them are essays. Essays paraphrase. Datasets get quoted.

What counts as original data when you’re not a research firm

Most small brands and agencies hear “original research” and assume it means commissioning a $40,000 panel study. It doesn’t. What an LLM needs is a number that is verifiable, sourceable, and absent from the rest of the open web. You almost certainly already have one.

Survey your customer list — even 80 responses produces a citable percentage. Pull your own platform metrics: response times, conversion rates, average ticket sizes, churn curves, support categories. Audit your industry’s public filings and publish the cleaned dataset with a methodology note. Run a 30-day teardown of pricing pages across your top 20 competitors and publish the spread. Scrape job boards in your vertical and chart the role-mix shift quarter over quarter. Any of these will produce numbers nobody else on the internet has packaged that exact way.

The format matters as much as the substance. The number has to live in a paragraph an LLM can lift cleanly — a single sentence, with the figure, the source (you), the time window, and the sample size. Bury it inside a slideshow or a downloadable PDF and you’ve made it invisible to the very engines you’re trying to feed.

What to do this week

First, find one number you already own and that nobody else has published. It can be small. “In 2026 our 412 surveyed restaurant clients reported a 28% jump in delivery-app fees year over year” is more citable than any opinion piece you’ve ever written.

Second, write the page around the number, not the other way around. The title states the finding. The first 30% of the page restates the finding with method, sample, and time window — the part LLMs disproportionately read. The middle explains why the number is what it is. The bottom links to the raw data or methodology.

Third, give the page a permanent home and never let it 404. Original-data pages accumulate citations over years. Treat the URL like infrastructure: clean slug, stable domain path, dated only inside the body.

Fourth, syndicate the number — not the article. Pitch the stat to industry newsletters, get it dropped into a Statista pull, push it onto Wikipedia where appropriate, mention it on a podcast transcript. Every additional surface that quotes your number, citing you as origin, strengthens the model’s confidence that you are the source.

Variant D — brand-targeted

If you’re a brand that wants to be the answer LLMs reach for (not just rank on Google), Paris Roussos has been engineering search visibility for 30 years and now runs done-for-you AI SEO. Flat-rate, no-fuss. Email parisroussos@gmail.com.

The backlink era rewarded brands that earned links. The AI search era rewards brands that earn citations — and the cheapest way to earn one is to publish a number nobody can paraphrase away.

For the last 18 months, almost every conversation about AI SEO has been about production — write longer pages, restructure the H2s, add stats, get on Reddit. All correct. All necessary. None of it is the bottleneck anymore.

The bottleneck is that almost nobody is measuring whether any of it is working.

AI Overviews now appear on roughly 48–50% of queries (BrightEdge, Feb 2026), AI-referred sessions grew 527% YoY through mid-2025, and AI traffic converts about 4.4× better than traditional organic. Yet most teams — agencies included — are still sending clients a Search Console screenshot and an Ahrefs traffic chart. Neither tool can tell you whether ChatGPT, Perplexity, Gemini, or Google AI Overviews currently cite you, mention you, or describe you accurately. That gap — measurement, not production — is the largest operational hole in the field right now, and whoever fills it first owns the conversation with the client.

The mechanic: what “AI visibility” actually decomposes into

Traditional SEO had one currency: position. AI search has four, and they don’t move together.

Citation share is the percentage of relevant prompts where your domain appears as a clicked source under an AI answer. ChatGPT cites Wikipedia 47.9% of the time. Perplexity leans Reddit 46.7%. Gemini stays on brand-owned sources about 52% of the time. Your citation share is engine-specific — averaging across them hides the truth.

Brand mention share is the percentage of relevant prompts where your brand is named in the answer text without necessarily being cited. This is the metric that maps to the +35% organic-clicks lift cited brands inside AIO see (Amsive). A brand can be mentioned without being clicked, and mentioned without being linked — both still move demand.

Representation accuracy is whether the AI describes you correctly when prompted. Wrong category, wrong founders, wrong pricing, outdated product description, ghosted-from-the-knowledge-graph — all of that is fixable, but only if you’re checking.

Share of voice is the head-to-head: across a fixed prompt set, what percent of citations or mentions go to you vs. each named competitor. This is the number that gets a CMO’s attention because it’s directly comparable to organic SOV reports they already read.

You need all four. A site can have rising citation share on Perplexity, falling mention share on ChatGPT, perfect representation accuracy in Gemini, and shrinking SOV against a single fast-moving competitor — all in the same week. The aggregated “AI visibility score” most vendors are selling collapses those four into one number and tells you nothing actionable.

What to do this week

1. Define your prompt set first, then pick tools. Pull 50–150 prompts directly from customer questions: sales-call transcripts, support tickets, your own Search Console long-tail queries, and the “People also ask” boxes around your top 20 keywords. This prompt set is the only ground truth in the entire stack. Refresh it quarterly.

2. Run the four engines manually for one week. Before you buy a tracker, hand-run your top 30 prompts against ChatGPT, Perplexity, Gemini, and Google AI Overviews. Log: cited? mentioned? described correctly? competitors named? Forty-five minutes a day for five days will teach you more than any dashboard.

3. Stand up automated tracking for the top 50. Profound, Otterly, AthenaHQ, Peec, and a handful of others now do this — pick one, pipe it into a sheet, and start a weekly snapshot. You don’t need real-time. You need a clean week-over-week trend on citation share, mention share, and SOV per engine.

4. Build the one chart that matters. A single stacked-bar showing your mention share vs. your three biggest competitors across the four engines. Update it weekly. That chart belongs in every client report from now on — and at the top of every internal review.

5. Set the dial for representation accuracy. Once a month, prompt each engine to “describe [your company]” and “list the top three companies that do [your category].” Fix anything wrong by editing the source of truth — usually Wikipedia/Wikidata, your About page, and your top three citation sources — not by trying to argue with the model.

You can’t optimize what you can’t see. The teams winning AI SEO right now aren’t the ones writing the most content — they’re the ones who can show a client, on a single page, whether last month’s work moved the share-of-voice line. The production playbook is already in the public domain. The measurement playbook is still wide open.

Variant B — direct, services-first

Need this done for you? Paris Roussos runs a flat-rate AI SEO service ($500–$1,500/mo per client, white-label for agencies) covering audits, schema and entity work, AI-visibility tracking, and content engineered to be cited by LLMs. Reach him at parisroussos@gmail.com.

If you don’t have the share-of-voice chart yet, you don’t have a program — you have a content calendar.

For a decade, the standard small-business content advice was “write tight.” 700 words, one keyword, hit publish. That formula still works fine for Google’s classic ten blue links. It is quietly catastrophic for AI search.

The hard data: pages above 20,000 characters of body text get 4.3× more AI citations than shorter pages. ChatGPT, Perplexity, Google’s AI Overviews, and Gemini are not reading your headline and guessing — they are retrieving chunks from inside long, structured documents and stitching answers together. Thin pages do not have enough surface area to be retrieved.

This is the inverse of what most founders and operators were taught about content, so it gets ignored. It shouldn’t.

The mechanic — why length is a retrieval signal

LLM-powered search does not work like ranking. A ranking system looks at the whole page and decides “this URL is #4.” A retrieval system slices your page into passages (usually 200–800 tokens each), embeds them into a vector space, and pulls the closest passages to the user’s prompt. A 700-word post gives the retriever maybe 2–3 candidate passages. A 4,000-word guide gives it 15–25.

More candidate passages means more chances of being the closest match to some sub-question the user asks. The model is not picking your “best” article. It is picking the best sentence-cluster for that exact prompt. Length increases your surface area.

There’s a second effect: long pages signal topical authority to embeddings models. When a single URL exhaustively covers a topic — definitions, sub-cases, exceptions, examples, comparisons — its semantic centroid sits closer to the topic’s centroid than a thin page does. The retriever’s similarity score climbs. This is why pillar-page strategies and topical clusters are now load-bearing, not optional.

The third effect is mundane but real: long pages get more internal links, more entity mentions, more anchored citations from your own site. That cluster of supporting signals reinforces the long page’s authority and pulls more retrieval traffic toward it.

“But long content is dead” — no, it’s the opposite

The “short, snackable content” advice came from a world where the goal was getting a single click off a SERP. In the AI search world, you are not competing for a click. You are competing to be the source the model quotes back. The mechanics flipped without most marketing teams noticing.

This is also why the +22% lift from adding statistics and the +37% lift from adding direct quotations show up so consistently — both reward longer, denser pages. You cannot squeeze that density into 500 words. You need room.

What to do this week

1. Audit your top 20 pages by traffic and pick the 5 with the strongest topical fit to your business. Run a character count. Anything under 8,000 characters is a candidate for expansion. Anything under 4,000 is barely a stub by AI search standards.

2. Expand by adding sub-topics, not filler. The goal is more retrievable passages, not more words. For each page, ask: “What are the five adjacent questions a buyer would ask after reading this?” Each one becomes an H2 with a tight 80–150 word answer beneath it. That’s how you cross the 20,000-character line without padding.

3. Add at least one original data point and one named expert quote per major section. Statistics and quotations are the two highest-lift content elements for AI visibility. They are also what the retriever extracts. If your page has neither, you are invisible to the citation layer regardless of length.

4. Keep the heading hierarchy strict. The retriever uses H2/H3 boundaries as passage breakpoints. Sloppy nesting collapses your passages into one giant chunk that nobody quotes. Treat your outline like API documentation.

5. Don’t merge short posts blindly. Two well-written 1,500-word articles on different sub-topics will out-cite one bloated 3,000-word amalgam. Length matters; coherence matters more. Expand each page on its own topic, don’t Frankenstein.

Paris Roussos has been doing SEO since 1996 (co-founded a Forbes Best of the Web–winning site back in the day) and now runs a white-label AI SEO practice for agencies and brands — flat-rate, $500–$1,500/mo per client. If your top-of-funnel traffic is leaking into ChatGPT and Perplexity and you want it back, email parisroussos@gmail.com.

The cheap-and-thin content era is over for AI search. The pages that get cited — and convert at the 4.4× rate AI-referred traffic delivers — are the long, well-structured, stat-rich ones. Build five of those before you publish another short post.