Cited but not recommended? Here's how to fix it

AI can read your page, put it in the sources, and still name someone else in the sentence the buyer remembers. Citation is the footnote. Recommendation is the shortlist. Track whether your name appears — not only whether your URL was used.

Search Console’s generative AI report landed in June 2026 with impressions for AI Overviews and AI Mode. No click column in that view yet.

A team screenshots the first AI impression. Then they read the answer.

Their URL is in the sources. Their name is not in the recommendation.

GPO’s August 2026 State of Search & AI put a number on it: when a brand’s own content was cited inside an AI Overview, that brand was still left out of the model’s actual recommendation 69% of the time. Google read the page, took a fact, and named a competitor in the line the buyer would repeat.

Rank is not the name. Another glossary post will not fix it. This is a recommendation problem — and it sits next to two mistakes we keep seeing in August 2026: treating the local map pack as AI discoverability, and treating a paid slot in AI Mode as citation lift.

If you still need extractable paragraphs, read how to get cited. If rank is green and visits are red, read you still rank number one. This page is for the team that already earns the footnote and still loses the name.

What is the difference between cited and recommended?

Citation is a footnote. The engine used your page — or a page that mentions you — as evidence for a claim. Your URL may sit in a sources drawer. A statistic from your blog may ground the summary. Pew Research Center (July 2025) found that only 1% of visits with an AI summary produced a click on the link inside that summary. Citation is often invisible to the buyer.

Google users are less likely to click on links when an AI summary appears — Pew, July 2025

Recommendation is the shortlist in the sentence. The answer names one to three options — vendors, studios, locations — as the thing to use, hire, or call. G2’s March 2026 buyer survey found 85% of B2B software buyers think more highly of a vendor the chatbot includes; 69% chose a different vendor than planned because of what the chatbot recommended; one in three bought from a vendor they had never heard of. The recommendation is what the buyer repeats. The citation usually is not.

Mention is weaker than both — your name in a sentence without a fact the engine had to retrieve. Easy to drop on the next refresh.

Citation is the source link. Recommendation is the name they remember.

We wrote the three-job split in SEO, AEO, and GEO. GEO that stops at citation stops before the buyer hears your name.

Why does AI cite you and recommend someone else?

Four patterns show up every time we run the sheet.

1. Your page supplied the fact. Their page supplied the pick.

Definitions, benchmarks, and how-it-works posts earn citations because the model needed a number or a clean sentence. Comparison and buyer-guide pages earn recommendations because the model needed a named option with a tradeoff. Princeton’s GEO study (KDD 2024) found citations, statistics, and quotations lift source visibility in generated answers. It did not promise the source becomes the recommended vendor. Your statistics paragraph can feed an Overview that names three agencies from G2, Reddit, or someone else’s comparison page.

2. The model trusts someone else’s page more than yours.

LinkedIn shows up often in professional-query citations. Reddit shows up often in local and software Overviews. G2 shapes software shortlists. A model that trusts Reddit, G2, or LinkedIn more than your site will cite you for a data point and recommend whoever else agrees off-site. B2B buyers start in ChatGPT is the pipeline version: the chatbot builds the list; the website confirms it.

3. Your page names the competitor — and the model picks them.

GPO’s August note describes the pattern plainly: the Overview reads your page, extracts a fact, and recommends competitors mentioned inside it. Your fair comparison post feeds the category. It does not feed your name unless the page says who you are for — not “we’re great,” but “unlike X, we do Y for teams that already have a site and a ticket export.”

4. Map pack and AI answer are different scoreboards.

GPO cited a study comparing AI recommendation rates to Google’s local 3-pack for the same brands: 35.9% appeared in the 3-pack; ChatGPT recommended those same brands 1.2% of the time; Gemini 11%; Perplexity 7.4%. Whitespark, July 2026 found AI Overviews on 68% of local business queries versus local packs on 39%. The Local SEO Data study (May 2026, 1,120 queries, 13 US cities) found 28.5% of businesses in the local pack were absent from AI Mode answers for the same search.

You can rank first on the map and never exist in the answer the customer trusts.

What should you measure instead of citation count?

If the only number going up is “we were in the sources,” you are polishing the footnote.

Track three things GPO groups under recommendation — not citation count alone:

Add Search Console’s generative AI report for Google-side citation exposure — impressions in Overviews and AI Mode, pages tab, device and country splits. Google announced the reports on 3 June 2026. That view shows impressions, not clicks. Pair it with the main Performance report for CTR on URLs that still earn traditional clicks.

Do not merge the two reports into one vanity line. AI impressions up, name still missing: the page supplied raw material for someone else’s shortlist.

For paid: GPO summarized research on 50,000+ commercial keywords in AI Mode — text ads on 29% of commercial searches, 54% where CPC exceeded $10. Only 11.5% of advertiser domains were also cited on the same queries; 2.3% of advertised URLs ranked organically for those terms. In AI Mode, paid, cited, and organic barely overlap. Ad spend does not buy a named slot in the answer. We run ads as one brief with the landing page on Meridian; we do not report ad impressions as GEO.

How do you fix cited-but-not-recommended?

This is the order we use. “Write more blog posts” usually recreates the problem.

1. Run the recommendation audit before you rewrite a paragraph

Write the twelve questions a buyer asks before they email you — in their words. Ask ChatGPT, Perplexity, Gemini, and Google AI Mode the same twelve on a fresh session. Record: named, accurate, cited URL, who else appeared, who won the recommendation.

We published the method in four models, twelve questions. The product is the sheet, not our August winners.

Cited on question 7 but absent on question 3? Build the page for question 3 — usually a comparison or a “best for [condition]” note, not another glossary.

2. Publish the comparison the model is already improvising

Models recommend from pages that say who to pick, when, and why — in a table. “Top 10 tools” listicles get paraphrased without you. “If you already have Shopify and need refund refusal in the widget, shortlist these two; if you need an owned agent on your help centre, shortlist these two” — that is a page worth citing and naming.

Include:

• A table with columns the buyer actually sorts by — price shape, handoff, who owns failure, ninety-day cost
• One statistic per section with year and publisher
• One quoted line with attribution — even if the quote is your own measurement rule from the number on Monday
• A plain “choose us when / not us when” block

A page that lists every option fairly and never says who it is for recommends no one — including you.

3. Make the same five facts true everywhere

Pick five facts that must match everywhere: company name spelling, primary service, price shape, geography, integration names, compliance boundary. Check the website, Google Business Profile if local, G2 or the relevant review profile, LinkedIn company page, and at least one post per month from a named person on LinkedIn.

LinkedIn’s 2026 guidance treats posts as readable text, not a link card. One specific post from a named person beats ten posts that only share a URL. If your blog says “agents on YourGPT and n8n” and your LinkedIn says “AI chatbots for everyone,” the model splits the difference and recommends neither.

For software: finish the G2 profile before you write another blog post. For local: fix NAP, categories, and hours across directories before you write another location page. Wrong postcode in an AI answer costs a visit; GPO flagged false location facts across 165 London businesses and 72,000 retailer questions — with no alert from the platforms.

4. Put the answer at the top of commercial URLs

Kevin Indig’s analysis of 1.2 million ChatGPT responses, cited in Search Engine Journal’s local AI piece, found 44.2% of citations from the first 30% of page content — the model often stops reading. Whitespark’s local work shows price and cost queries trigger Overviews more than 80% of the time. Most businesses still refuse to answer them.

On /services, pricing, and comparison URLs, put:

• A 40–60 word direct answer — claim, number, source, who it is not for
• A table the model can lift without interpreting prose
• An FAQ where the first sentence is the answer

We showed the before/after on a service block in how to get cited. Recommendation work adds who you are for — in plain language — on top of the extract.

5. Get one true sentence off your site

Omniscient Digital’s study of 23,000+ citations, cited in the same SEJ piece, found owned content accounted for roughly 23% of citations on branded queries — 77% off-page. The model names you after it sees the same fact twice.

One line of effort: a case outcome specific enough to repeat — “support agent on dirty tickets, same morning volume, handoff only on exceptions” — on /work/aurelia, in a customer quote if you have permission, and once in a place a model already reads (review, podcast, trade press, an honest answer in a Reddit thread). Not astroturf. One true sentence in two places beats five blog posts that say the same adjective.

6. Split the quarterly board into four columns

Kill pages that win neither a click nor a citation nor a recommendation. Refresh pages that win citation but not name — add a comparison table, a fit block, off-site alignment. Build new URLs only where the audit shows a missing question, not a missing keyword.

What does this look like on a real studio page?

You run marketing for brands with existing sites. ChatGPT cites your journal on zero-click statistics but names three other agencies for “who should run Meta and Google.”

What we see: cited in the footnote, absent from the shortlist. The statistics post feeds the Overview. It does not name you. There is no comparison URL that says when Simplex is the fit.

What we would do:

1. Publish or rewrite a comparison: “Who should run Meta and Google when the site already exists” — in-house vs freelancer vs studio vs media-only shop; ninety-day cost shape; choose a studio when ads, landing page, and Monday inquiry must be one brief.
2. Add a 52-word extract at the top of /services#marketing with Ahrefs 58% (Dec 2025), Pew 8% vs 15% (July 2025), and the fit line — same method as how to get cited.
3. Post once on LinkedIn from a named partner with the fit paragraph. Not a link card. The facts must match the site word for word.
4. Rerun prompts 4, 6, and 12 from four models, twelve questions in thirty days.

What we would not do: ten more posts defining AI search. A new llms.txt entry with no HTML changes. Opting out of AI Overviews — that removes the footnote. It does not put your name in the sentence.

What should you refuse to do?

Do not report citation impressions as pipeline. Generative AI impressions without recommendation tracking decorate the deck.

Do not treat map position as AI recommendation. Run GPO’s local prompt set — discovery, comparison, trust, logistics — across ChatGPT, Gemini, Overviews, and Perplexity. Score mention rate and factual accuracy, not map rank alone.

Do not buy ads to “boost GEO.” Paid and organic citation barely overlap on the same query.

Do not fair-list five competitors without saying who you are for. That page helps the category. It recommends someone else.

Do not chase keyword stuffing. Princeton GEO-bench: keyword stuffing scored 17.7 on visibility vs 19.3 baseline — worse than doing nothing.

What belongs on the site after the fix?

The name has to hold up when they open the site. Put these where a stranger and a model can both read them:

• /work — four cases with constraints, not adjectives
• A pricing or engagement page a model can read without guessing — see llms.txt and the file stack
• A contact line that is a brief, not a demo trap: “Send the twelve questions you ask ChatGPT before you hire, and the URLs you think should answer them. We will write back with which prompts cite you, which recommend you, and what we would change on one page.”

FAQ

If AI cites my page, doesn’t that mean visibility is working?

No — not if your name is missing. Citation means your page supplied a fact. The buyer may never click. GPO’s August 2026 research: brands were left out of the recommendation 69% of the time when their own content was cited in an AI Overview. Visibility in the footnote is not visibility on the shortlist. Track both.

How is a recommendation different from a mention?

A mention is your name in a sentence nobody repeats. A recommendation is your name as an option to choose — usually with a reason or a comparison. G2’s 2026 buyer work found 69% chose a different vendor than planned based on chatbot recommendations. A mention without a reason drops on refresh.

Does map-pack #1 get you named in AI answers?

Not often. GPO cited 35.9% 3-pack appearance vs 1.2% ChatGPT recommendation for the same brands. Whitespark (July 2026): AI Overviews on 68% of local queries vs packs on 39%. Same city, same query, different winner. Do not read map rank as AI rank.

Will Google Search Console tell you if you were recommended?

No. The generative AI performance report shows how often your URLs appeared in Overviews and AI Mode — not recommendation share, not query text, not clicks in that view. Use Search Console for Google citation exposure. Use the twelve-prompt sheet for recommendation share across engines.

Do AI Overviews ads help you get cited or recommended organically?

No. Ads and organic citation barely overlap on the same query. GPO summarized research: 11.5% of advertiser domains were also cited on the same AI Mode queries; 2.3% of advertised URLs ranked organically for those terms. Judge ads on conversion. Build the name with a comparison page and one off-site proof line.

Which page type should you write first?

A comparison or buyer-guide URL with a decision table, a named statistic, a quoted rule, and a “choose us when / not us when” block. Informational posts earn citations. Comparison pages earn recommendations. Start with the prompt where you are cited but not named.

Should you opt out of AI Overviews if competitors get recommended instead?

Usually no. Opt-out removes your URLs from generative features without changing traditional rankings. Fix the name first — fit content, comparison page, off-site alignment. Opt-out is for strategic exclusion, not wounded pride.

How often should you rerun the twelve buyer prompts?

Every 30 days. Same wording, same four engines, fresh session. If citation improved but recommendation did not, rewrite the comparison block and align off-site facts — not another definition post.

Send the twelve questions you ask before you hire — or that your buyers ask before they hire you — plus the URL you think should win the recommendation. We will write back with which prompts cite you, which name a competitor, and the one page we would rewrite first. Contact the studio.