Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- AI Search
- Link Building
Does publishing original research lift AI citations more than it lifts backlinks, or have you seen both move together every time?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
They do not reliably move together, and public evidence cannot yet tell you which one original research lifts more. Both come from the same place, other people finding your data worth repeating, but they are credited differently. A backlink accrues to your domain. An AI citation goes to whichever page the system retrieved and chose, and that can be a news article about your study rather than the study itself.
So I would expect them to rise together at the campaign level and come apart at the page level. Plan research for both, track them as separate outcomes against separate baselines, and build the asset so that when an answer uses your number, the citation has a fair chance of landing on your page.
Why links and citations can rise together
The shared path is coverage. A study that journalists and bloggers pick up earns links and mentions in the same weeks, and links still feed rankings. Rankings matter for AI answers built on search results. Google’s guide to optimising for its generative AI features (May 2026) says foundational SEO still applies to them. OpenAI’s crawler documentation separates OAI-SearchBot, which surfaces sites in ChatGPT’s search features, from GPTBot, which collects training data. A separate ChatGPT search help page says ChatGPT search sometimes partners with third-party search providers, so rankings outside OpenAI’s own crawl can feed it too. A study that earns links and ranks is more likely to be in the pool these systems retrieve from.
But ranking is a loose predictor of citation. Ahrefs reported the share of AI Overview cited pages that also rank in Google’s top 10 falling from 76% in July 2025 to 38% in March 2026, with the caveat that its detection method changed between the two datasets. BrightEdge measured 17% overlap with a different method. Whichever figure is closer to the truth, a large share of cited pages sit outside the top 10, which leaves room for citations to move while the link profile stands still.
Four reasons they come apart
The citation goes to the coverage
A September 2025 preprint by Chen, Wang, Chen and Koudas compared AI search with Google and found “a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content”, in contrast to Google’s more balanced mix. If that pattern holds in your topic, the page most likely to be cited for your statistic is a publisher’s write-up of it. Your link count rises; your own citation count may not. Whether those answers name you as the source is a separate outcome, and one worth tracking. The same study found that AI search services differ from each other in domain diversity, freshness and sensitivity to phrasing, so “AI citations” is several numbers, not one.
Specific numbers help a page get picked, once it is already a candidate
The GEO paper by Aggarwal and colleagues, presented at KDD 2024, rewrote sources in a benchmark and measured how visible they became in generated answers. Its best methods, “Cite Sources, Quotation Addition, and Statistics Addition”, achieved a relative improvement of 30 to 40% on its Position-Adjusted Word Count metric, while keyword stuffing offered “little to no improvement”. The effect was not uniform by rank: the Cite Sources method raised visibility for sources ranked fifth by 115.1%, while visibility for top-ranked sources fell by 30.3% on average.
Two limits matter for your question. The engine chose among the top five Google results for each query, with gpt-3.5-turbo writing the answers, so the study measures selection among pages already retrieved, not whether a page gets retrieved in the first place. And it tested adding statistics, not whether they were original. Original research gives you numbers nobody else has, which fits the mechanism, but originality itself was not tested.
Links arrive in a news window; citations follow question demand
Links cluster around launch and outreach. Citations accrue when people ask questions your data answers, which can continue for years for a useful benchmark (typical costs, rates, timings) and stop within weeks for a novelty statistic. A plain annual benchmark can out-earn a headline-grabbing survey in citations while losing to it on links.
Format decides whether the number is quotable
Journalists can link a PDF or a gated report if the headline is strong. A retrieval system can only quote what it can extract, so findings that live only in a PDF, a chart image or a script-rendered widget give it less to work with. Google’s guide says you do not need special markup, artificial chunking or AI-specific rewording for its AI features, so this is plain accessibility: the key numbers in HTML text, with the definition, sample and date beside them.
Measure them as separate outcomes
Outcome Credited to Where to measure Watch out for Backlinks Your domain Link index tools, filtered to the findings URL Syndicated copies inflating counts Unlinked mentions Your brand or dataset name Mention monitoring Name collisions with other brands Google AI feature visibility for your page Your URL Search Console’s “Search generative AI performance” report, filtered by page; available to all eligible sites since 31 August 2026 Impressions only: no clicks, CTR or queries Copilot and Bing AI citations Your URL Bing Webmaster Tools AI Performance report (launched in public preview in February 2026): citations, cited pages and a sample of grounding queries No click or traffic data Answers that cite coverage and name you The publisher’s URL Prompt sampling that records cited URLs and whether you are named Invisible in any page-level report AI referral visits Your site GA4’s “AI Assistant” channel group (added May 2026) Visits without a referrer land in Direct; practitioners report AI Mode clicks arriving as google / organic How to test it on your next study
“More than” only means something against a baseline, so define both outcomes before launch and compare each with its own history or with a control asset.
- Write the prompt set first. List 20 to 50 questions your data should answer, phrased the way customers, journalists and analysts would ask them, plus a few control questions on the same topic that your data does not answer.
- Take a baseline. Run each prompt several times per platform, because answers vary between runs, and record the cited URLs and whether your brand is named. Record referring domains and mentions for the same period.
- Stagger the release. Publish the findings page and link it internally, but hold outreach until URL Inspection shows it indexed and you have at least two rounds of prompt samples. Then run outreach. The quiet phase shows what the data does on its own; the outreach phase shows what coverage adds.
- Run the outreach phase until the coverage itself is indexed and sampled at least twice. That ties the window to events you can observe rather than to a fixed number of weeks.
- Keep a control. A comparable non-research asset published in the same period, or your previous campaign’s numbers, tells you whether a citation lift is yours or a platform-wide shift.
- Log confounders. Model and product changes can shift citation behaviour without notice, a core update can reshuffle the rankings that retrieval draws on, and news cycles change what people ask.
Result pattern Likely reading What I would do next Links up; your own citations flat; coverage cited and naming you The earned-media pattern Count named citations as a win; ask coverage to link the findings page; make the page easier to quote Citations rise in the quiet phase, before any links The data answers questions people already ask Invest in update cadence on the same URL; treat links as a bonus Links up; nobody cited for your prompt set The statistic is not what people ask, or the prompts miss the real questions Revisit the research questions before the outreach plan Both flat Weak newsworthiness and weak question demand Test a different kind of data before scaling the programme Making the citation more likely to land on you
- One canonical findings page in HTML, on a stable URL you update each edition, with past editions archived and linked from it.
- Key numbers written as text, with the definition, sample size, method and fieldwork dates next to them.
- A distinctive name for the dataset, so mentions and AI answers can attribute it to you even without a link.
- The findings URL, not a PDF or the homepage, in every pitch and press release.
- No manufactured mentions. Google’s guide says seeking inauthentic mentions is not helpful, and seeded mentions distort the very metric you are trying to read.
Where I would land
I would fund original research against two separate tests: will people cover it, and will people keep asking the questions it answers? Links mostly tell you about the first and citations about the second, and the two overlap less than a single campaign dashboard suggests. Report them separately, count answers that name you through someone else’s coverage, and judge the research on the combination rather than expecting one number to predict the other.
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.