Business Growth

Competitive Intelligence in an Agentic Market: Monitoring When Agents Do the Shopping

Competitive intelligence used to mean watching rankings, ads, and pitch decks. In an agentic market, the decisive comparison increasingly happens inside an AI conversation you never see. Gartner projects machine customers, AI systems that research and buy on humans' behalf, will be involved in or influence a cumulative 30 trillion dollars in purchases by 2030, and surveyed CEOs expect them to drive up to 20% of revenue by then (Gartner, 2024). The transaction rails are already shipping: OpenAI and Stripe launched the Agentic Commerce Protocol and Instant Checkout in ChatGPT in 2025 (OpenAI, 2025). For service firms, the question shifts from 'where do we rank' to 'what do agents say about us, against whom, and why.' This article builds that monitoring discipline.

Joshua Agonya Pi'Rwot

By Joshua Agonya Pi'Rwot

Founder, Business Growth Accelerator

Executive summary

When AI agents shortlist vendors, competitive position is decided inside model outputs you cannot see. The evidence on machine customers, plus a monitoring system for tracking how assistants describe you and your rivals.

Section 1

The five challenges at a glance

Moving CI into the agentic era surfaces five problems classic competitive monitoring never had to solve. The evaluation surface is invisible, recommendations form inside model outputs, not public SERPs; machines weigh different evidence than human buyers do; answers are volatile across model updates; rivals are now optimizing for citations deliberately, making visibility zero-sum; and the tooling for tracking any of this is young and noisy. The table maps each to its root cause, the most exposed firms, and the controlling evidence. The strategic point underneath all five rows: in agent-mediated markets, your competitive position is partly an artifact of training data and retrieved sources, which means it can be audited, diagnosed, and deliberately improved.

Section 2

Challenge one: the buying evaluation has moved somewhere you cannot see

Gartner's machine customer research describes a three-phase evolution: bound customers executing human-set rules, adaptable customers choosing within human constraints, and autonomous customers acting with broad discretion (Gartner, 2024). Service procurement sits early on that curve, but the research phase, where shortlists form, has already shifted. ChatGPT passed 800 million weekly users in late 2025 (OpenAI, 2025), and Pew found that when an AI summary answers a query, users click traditional results in only 8% of cases (Pew Research Center, 2025). A buyer who asks an assistant to 'compare the top three RevOps consultancies for a 40-person SaaS firm' receives a synthesized verdict assembled from sources you may not know exist. Nothing in your analytics records that comparison; you see only its downstream effects, a branded search, an inquiry, or silence. The transaction layer is following: Instant Checkout in ChatGPT launched with Etsy sellers and announced support for over a million Shopify merchants, with Salesforce backing the Agentic Commerce Protocol weeks later (OpenAI, 2025; Salesforce, 2025). Services will lag products on automated purchase, but the shortlist, the part of the funnel where most competitive battles are decided, is already agent-mediated. CI that monitors only visible surfaces is auditing a stage the decision has partly left.

Section 3

Challenge two: machines weigh different evidence than humans do

Human buyers respond to chemistry, brand aesthetics, and referrals. Generative engines respond to extractable, verifiable, well-attributed claims. The Princeton-led GEO research quantified this: adding statistics, quotations, and source citations lifted content visibility in generative answers by up to 40%, while traditional persuasion signals did little (Aggarwal et al., 2024). Engines also cross-reference, Pew found 88% of AI summaries drew on three or more sources (Pew Research Center, 2025), so a competitor whose claims are consistent across their site, directories, review platforms, and press coverage presents a more confident citation target than a firm whose footprint is contradictory or thin. This redraws competitive advantage in ways operators underestimate. A rival with mediocre delivery but excellent machine-legible evidence, published benchmarks, structured data, specific named outcomes, can outperform a genuinely better firm inside agent recommendations. There is also a structural difference in market timing: Ehrenberg-Bass research shows about 95% of buyers are out-of-market at any moment (Ehrenberg-Bass Institute, 2021), and human mental availability is built over years. An agent, by contrast, assembles its consideration set fresh at query time. That cuts both ways, incumbent brand memory matters less inside the model's synthesis, which means challengers can win agent shortlists they could never win in human recall, and incumbents can quietly lose positions they assume are safe.

Section 4

Challenge three: the answers keep moving and the tools are young

Even firms that grasp the shift hit a measurement wall: agentic visibility is volatile and the instrumentation is immature. Semrush documented AI Overview trigger rates swinging from 6.49% of queries in January 2025 to 24.61% in July and back to 15.69% by November, a threefold oscillation in a single year on one surface alone (Semrush, 2025). Underneath presentation changes, model versions, retrieval pipelines, and source weighting all shift without notice, and answers vary across phrasings, sessions, and user contexts. A single spot-check, 'I asked ChatGPT and we came up first', is anecdote, not intelligence. The vendor ecosystem responding to this (AI visibility trackers, share-of-answer dashboards, including offerings from Semrush and newer specialists) is genuinely useful but young: methodologies differ, sampling is partial, and no tool observes private conversations directly, flag any vendor metrics as directional estimates rather than ground truth. The practical consequence is that agentic CI must be built on repeated, standardized sampling you control: the same prompt set, the same assistants, the same scoring rubric, run on a fixed cadence so that movement in the data reflects market movement rather than measurement noise. Cloudflare's finding that user-driven AI crawling grew 15-fold in 2025 confirms the underlying activity is real and growing (Cloudflare, 2025); the job is building instruments that see it.

Section 5

Innovative solutions

Five practices form the working toolkit. First, the prompt panel: 25-40 standardized queries reflecting real buyer language, category comparisons, 'best X for Y' questions, problem-first phrasings, run monthly across ChatGPT, Claude, Gemini, and Perplexity. Score each response for whether you appear, what is claimed, sentiment, and which competitors share the answer. Second, citation tracing: where assistants cite sources, log them. Those third-party pages, review sites, directories, comparison posts, news mentions, are the upstream levers of your downstream visibility; a single well-maintained profile on a frequently cited directory can outweigh ten blog posts. Third, competitor instrumentation: monitor rivals' schema deployment, llms.txt files, published statistics, and review velocity. A competitor suddenly shipping structured data and citable benchmarks is telling you their strategy (Aggarwal et al., 2024). Fourth, the correction protocol: when assistants state something false about your firm, wrong pricing, dead services, misattributed work, fix the source material the models retrieve from, since you cannot edit the model: update profiles, correct directories, publish authoritative pages addressing the error directly. Fifth, win the citable-facts war: publish specific, verifiable, dated claims, benchmark data, named outcomes with numbers, original research, because the GEO evidence shows these are precisely what engines select for (Aggarwal et al., 2024). Together these convert an invisible competitive arena into a measured one for a few hours of effort per month.

Section 6

Solution framework

Run agentic CI as a four-stage loop: Observe, Diagnose, Act, Verify. Observe is the monthly prompt panel plus citation logging, output: a citation share number (the percentage of panel queries where you appear), claim-accuracy rate, and a competitor co-mention map. Diagnose asks why the numbers are what they are: trace which sources feed answers about your category, identify where you are absent or misrepresented, and classify gaps as content gaps (the fact does not exist anywhere), distribution gaps (it exists but not on cited sources), or consistency gaps (sources contradict each other). Decompose competitor wins the same way, when a rival owns an answer, identify the specific pages earning it. Act assigns each gap an owner and a fix: publish, syndicate, or reconcile. Verify re-runs the panel and attributes movement honestly, remembering surface volatility means trends matter over quarters, not single cycles (Semrush, 2025). Govern the loop with three KPIs reported alongside traditional pipeline metrics: citation share, accuracy rate, and source coverage (the share of frequently cited category sources where you are present and correct). Keep the human-market loop running in parallel, agents mediate research, but humans still award service contracts, and the 95:5 logic of brand building remains intact (Ehrenberg-Bass Institute, 2021). The firms that win will run both loops without confusing them.

Section 7

Evidence-based action plan

Days 1-30: build the instrument. Draft your prompt panel from real buyer language, mine inquiry emails and sales call notes for actual phrasings. Run the baseline across four assistants, score it, and log every cited source into a category source map. You now know your citation share, your accuracy rate, and which third-party pages govern your market's answers. Days 31-60: fix the cheapest gaps first. Correct false claims at their sources, complete and reconcile every major directory and review profile, and ship structured data on your own site so your claims are machine-verifiable (Aggarwal et al., 2024). Begin competitor instrumentation: archive how assistants describe your top three rivals so future shifts are detectable. Days 61-90: go on offense. Publish two citable assets, original benchmark data, a priced service comparison, dated research, engineered for extraction. Re-run the panel; expect noisy movement and judge direction, not single deltas (Semrush, 2025). Ongoing: monthly panels, quarterly deep diagnosis, and a standing agenda item in your growth review: what do the machines say about us this month, and why? Budget honestly, this is hours per month, not a platform purchase, and treat vendor visibility tools as accelerants once your manual loop proves which metrics matter. In agent-mediated markets, the firm that measures the conversation first usually ends up steering it. For adjacent evidence in this pillar, see [Data Moats for Small Firms: Proprietary Data as Defensibility in the Agent Era](/blog/growth-data-moats-small-firms-agent-era) and [The Post-Platform Risk: Platform Dependence Research and the Case for Protocol-Native Businesses](/blog/growth-post-platform-risk-protocol-native).

FAQ

Direct answers for operators.

What are machine customers and do they matter for service businesses?

Machine customers are nonhuman economic actors, AI systems that research, select, and pay for goods or services. Gartner projects they will be involved in or influence 30 trillion dollars in purchases by 2030, with CEOs expecting up to 20% of revenue from them (Gartner, 2024). For services, full machine purchasing is further out, but agent-mediated research and shortlisting is already routine, which is where competitive position is decided.

How do I find out what AI assistants say about my firm?

Run a standardized prompt panel: 25-40 realistic buyer questions across ChatGPT, Claude, Gemini, and Perplexity on a monthly cadence, scoring citation presence, claim accuracy, sentiment, and competitor co-mentions. Single spot-checks mislead because answers vary across phrasing and model updates, Semrush documented AI Overview behavior swinging threefold within 2025 (Semrush, 2025). Repeated, consistent sampling is what separates intelligence from anecdote.

Can I correct false information an AI assistant states about my business?

Not directly, you cannot edit the model. You fix the sources it retrieves from. Trace cited pages, correct directory listings and review profiles, reconcile contradictions across your web presence, and publish an authoritative page addressing the specific error. Because most AI summaries synthesize three or more sources (Pew Research Center, 2025), consistency across your footprint is what shifts the answer over subsequent cycles.

What should a service firm publish to win agent recommendations?

Citable, verifiable specifics. The Princeton GEO research found adding statistics, quotations, and cited sources lifted generative-answer visibility by up to 40% (Aggarwal et al., 2024). For a service firm that means original benchmark data, named outcomes with numbers, transparent pricing logic, and structured data that lets machines verify claims. Vague positioning language is invisible to this channel; dated, specific facts are its currency.

Joshua Agonya Pi'Rwot

Written by

Joshua Agonya Pi'Rwot

Founder, Business Growth Accelerator · Country Director, AVODA Group Uganda · EMBA

Joshua helps service-business operators turn scattered marketing into a clear path from first attention to booked call. He is Founder of Business Growth Accelerator and Country Director of AVODA Group Uganda.