Generative engine optimisation: evidence before tactics

A GEO brief should name its intended result. A source cited in an explanation, a brand included in a shortlist and a visit to a service page are different outcomes. Before commissioning work, agree which one you will observe and why it matters.

The label does not tell you which systems a supplier has tested, what access they have or how they will measure change. Ask for those details in ordinary language. A useful proposal identifies the pages, the work and the limits.

For Google's AI Overviews and AI Mode, a supporting page needs to be indexed and eligible to show a search snippet. Check Googlebot access, page indexing directives and the rendered text. Make the page reachable through relevant internal links. Structured data should describe the content visitors can read. Google specifies no special AI schema or AI text file for these features.

Search Console also has a Search generative AI control under Settings. Check the effective choice for the domain and any relevant URL prefix property, including inherited settings. Inclusion is the default. Exclusion prevents the site's links and content from appearing in the covered AI features while leaving ordinary Search ranking outside this control.

OpenAI distinguishes its search crawler from its training crawler. Allowing training is a separate choice from allowing search access. Check the search crawler's robots rules and whether the hosting firewall allows its published IP ranges. A successful request made with a copied crawler name alone cannot establish that the real crawler has access.

Treat crawl access as eligibility. A clear answer still needs useful evidence, and each platform decides which sources to include. Resolve a recorded access problem before commissioning another article about the same question.

Use a record another person could review: exact question, platform, date, market, relevant settings and linked sources. Save the full response, including cases where the company is absent. Record a branded question separately from a request that does not name the business.

Repeat a manageable sample under comparable conditions. Describe it as an observation of that sample. Avoid presenting a few favourable screenshots as a share of all possible answers, or as an explanation of the platform's internal weighting.

Google's Generative AI performance report shows impressions for AI Overviews and AI Mode, with page, country, date and device breakdowns. The report may be absent when the site has insufficient impressions. The AI report draws on data included in the Web search report, so adding the totals would count some impressions twice. Assess recorded citations, visits and qualified enquiries through their own evidence.

Bing Webmaster Tools offers a separate AI Performance report for citation activity across supported Microsoft AI experiences, including Copilot and Bing summaries. Review total citations, the cited page URLs and the available sample of grounding queries over a stated period. The report measures recorded source use; ranking, placement within an answer and a recommendation to hire the publisher remain separate questions.

For a business serving several countries, use each report's available country filters and retain the selected dates, pages and device scope. Where a report lacks a country breakdown, leave the market unassigned; a country name in the page address or a grounding query alone is insufficient evidence of the reader's location. Keep platform citations, Google AI impressions, referral visits and qualified enquiries in separate columns.

Ask for the original shared conversation, export or complete capture showing the question, response and source links. Retain the run date, timezone, displayed model when available, search setting, market and prior conversation context. Record whether the question named the business or supplied its URL. Label missing details as unknown.

Open the exact cited URL and check what the answer used it to support. A source for advice on choosing an agency differs from a recommendation to hire that agency. A response to one question also provides evidence only for that question; a broader study title cannot expand its scope.

OpenAI explains that the Sources panel can include cited sources alongside other relevant links. Record where the page appears: a citation in the answer, a link only in Sources or an unlinked business mention. Keep those categories separate when calculating citation counts. A panel listing alone cannot establish that the response used the page to support a claim.

For a repeated study, request the complete question list, run counts, unsuccessful runs, selection rules and how citations were extracted. Report the number of qualifying responses against the total completed sample, with missing records identified separately. Preserve the evidence behind each counted result.

Publish a finding only within those boundaries: the observer, date, exact question, linked page and what the response actually said. A recorded citation gives neither a platform endorsement nor a promise of future visibility. A badge that removes these details gives a buyer less useful evidence than the original record.

Record the exact error, check it against a reliable source and correct any inaccurate page you control. Save the question, complete answer, cited URLs, date and product settings before making changes. A missing business mention is different from an incorrect business fact.

For a service business, check the trading name, responsible person, service scope, delivery location and contact route against the current website. If the answer cites an outdated page, improve that page and its relevant internal links. For an error on another publisher's page, record the source and the correction needed. Use the product's feedback option where available, and retain a copy of any report you submit.

Repeat the same question under comparable conditions and keep both observations. Mark a correction as observed only when a saved response shows it. A corrected source page or a submitted feedback report cannot establish that an assistant has updated its answer. Keep unavailable observations separate from completed checks, and record other changes that could affect the comparison.

Aggarwal and colleagues introduced a benchmark and evaluated methods for changing source visibility in generated responses. The paper reports gains of up to 40% on its tested visibility measures and says effectiveness varied across domains.

That figure is neither an average uplift promised to every site nor a 40% increase in enquiries. It concerns the benchmark and systems in the study. The abstract alone also cannot support a universal claim that writing polish is ineffective or that smaller sites always benefit most.

For your own review, turn a proposed technique into an editorial question: does this addition help a buyer check the answer? A relevant project detail may do that. A statistic added solely to make the page look authoritative may not.

Consider an illustrative remote service provider serving UK clients. A page that lists strategy, content and design leaves the buyer with a broad menu. The provider could explain the first decision, the materials required and which outputs belong to the proposed engagement.

Add a documented example with a clear boundary: what was delivered, which outcome was measured and which was not. Keep an illustrative scenario visibly labelled. This creates useful material for a buying conversation even if no assistant cites the page.

Google's helpful-content guidance cautions against changing dates without substantive changes. Record an editorial update when the work warrants it; a new calendar date alone is not new evidence.

Request permission before publishing client material. Keep the client's wording and the context of any measured result. A testimonial about working together cannot establish an unmeasured commercial outcome.

When a third party describes the practice, check factual errors and let that source retain its own voice. Requiring everyone to repeat one exact sentence would obscure the difference between independent comment and coordinated promotion.

Choose communities because you can contribute something useful to their readers. A post does not establish whether a future system will train on it, cite it or recommend its author.

Before accepting a promise, ask to see its comparison: what changed, what stayed the same, how often the result was observed and whether the evidence concerns your market. A demonstration should survive questions about unfavourable runs and missing data.

Avoid invented numbers, fabricated reviews and hidden instructions intended to steer an answer. These make the record less trustworthy for a buyer. Their existence does not establish that they are effective techniques.

Google says its generative Search features need no special schema or AI text file. Read a platform's own guidance before paying for a claimed technical requirement, and keep that guidance scoped to the platform that issued it.

Days one to fifteen: choose a narrow question set, establish the baseline and identify the most useful page to improve. Include a technical access check when the page cannot be fetched or indexed.

Days sixteen to sixty: improve that page with verified evidence and a clearer explanation. Keep a change log. Decide whether another question belongs on the same page before expanding the library.

Days sixty one to ninety: repeat the observations, review any visits and suitable enquiries, and decide what the evidence supports doing next. These are suggested work windows, not a forecast of when an assistant will cite the business.